【摘要】 为解决挖掘机支重轮垂直载荷分布不均衡问题，构建基于ADAMS多体动力学与贝叶斯正则化神经网络的耦合框架，用于优化支重轮垂直载荷的均衡分布。首先，通过ADAMS建立整机匀速直线行走虚拟样机，提取9个支重轮的垂直载荷基线数据。随后，采用DOE全因子设计，在±70 mm极限位置对支重轮2～8进行2~7次组合仿真，构建了129组高置信度训练样本。基于这些数据，建立了贝叶斯正则化神经网络，实现了位置-载荷映射的毫秒级预测。优化后，载荷分布的标准偏差从3 306 N降至1 370 N，降幅达58.6%。神经网络预测值与ADAMS验证结果的最大相对误差不超过4%，满足工程精度要求。在计算效率方面，传统穷举法需完成1.7×10~8次仿真，预估耗时6 834万h，而本框架全程仅需3 h，加速比达7个数量级。更多还原

【Abstract】 To address the uneven distribution of vertical loads on excavator track rollers,a coupled framework based on ADAMS multibody dynamics and a Bayesian Neural Network is constructed to optimize the balanced distribution of vertical loads on track rollers.First,a virtual prototype of the entire machine travelling in a straight line at a constant speed is built using ADAMS,from which baseline data of vertical loads on the nine track rollers are extracted.Subsequently,a full factorial DOE is performed to conduct27 combined simulations on track rollers2 to 8 at extreme positions of±70 mm,generating 129 sets of high-confidence training samples.Based on these data,a Bayesian Neural Network is established to achieve millisecond-level prediction of position-load mapping.After optimization,the standard deviation of load distribution drops from 3 306 N to 1 370 N,with a reduction of 58.6%.The maximum relative error between the neural network prediction value and the ADAMS verification result is less than 4%,meeting the engineering accuracy requirements.In terms of computational efficiency,the traditional exhaustive method requires1.7x108 simulations,estimated to take68.34 million hours,while this framework only takes 3 hours for the entire process,with a speedup of 7 orders of magnitude.更多还原