融合多头注意力机制的货车载重估计模型

Weight estimation model for trucks integrating multi-head attention mechanism

为了提高实时货车载重估计的便捷性和精准度,辅助开展大范围低等级公路网的货车运输过程中载重的实时监管,本文从货车动静态信息的交互效应出发,提出了一种融合多头注意力机制的货车载重估计模型(Mix-MAN)。首先,在模型中引入了多头注意力机制,增强网络对运动学时序特征的提取能力;其次,利用堆叠自编码器捕获货车的静态特征;最后,设计了一种特征融合结构,融合提取动态特征和静态特征,建立输入特征与货车载重之间的非线性映射关系,得到货车载重估计结果。试验结果表明:与未考虑货车静态信息的MAN模型对比,Mix-MAN的平均绝对值误差减小了6%,均方根误差减小了5%,平均绝对百分比误差减小了0.5%。本文模型可为我国公路货物运输监管、道路养护等方面提供技术支持。

In order to improve the real-time truck load estimation of convenience and accuracy,and to help to develop real-time monitoring for truck loads in a large-scale low-grade highway networks,this paper proposed a weight estimation model(Mix-MAN)for trucks integrating the multi-head attention mechanism based on the interaction effect of dynamic and static information of trucks.Firstly,multihead attention was introduced into the model to enhance the network′s ability to extract kinematic time series features;secondly,a stacked auto-encoder was used to capture the static features of trucks;finally,a feature fusion structure was designed to extract dynamic features and static features,establish the nonlinear mapping relationship between input features and weight estimation,and then to obtain the final weight estimation result of trucks.The experimental results show that compared with the MAN model without considering static information of truck,the mean absolute value error of Mix-MAN is reduced by 6%,root mean square error is reduced by 5%,mean absolute percentage error is reduced by 0.5%.The model in this paper can provide technical support for the cargo transport supervision of highway and road maintenance.