针对AP聚类RBF神经网络在车辆动态称重应用中精度偏低问题,提出按一定步长,迭代增加偏向参数,以RBF神经网络测试误差为评价指标最终确定偏向参数的改进算法,使RBF神经网络获得合适的隐含层节点数;提出对测试样本进行归类插值分析,利用与测试样本至类代表点间距离最接近的两个训练样本实际连接权值,使RBF神经网络连接权值随测试样本改变的自适应功能.在车速10 km/h~50 km/h,温度16 ℃~29 ℃条件下,对5种不同载重车辆进行工程实测试验,构建车辆动态称重RBF神经网络模型,进行500次循环测试.试验表明,基于AP聚类RBF神经网络的改进算法使称重误差均值控制在0.06%以内,最大实时性均值为0.0223,能有效满足实际工程应用要求.%An improved algorithm is proposed to solve the lower application precision of radical basis function (RBF)neural network based on affinity propagation(AP)clustering on the vehicle weigh-in-motion. This algorithm takes RBF neural network test error as the criteria to increase iteratively the preference which is obtained with fixed step length. In this way,appropriate hidden layer nodes are obtained. Classification and interpolation analysis of test sample is carried out based on the actual connection weight of two training samples which are nearest between the exemplar and the test sample,making the connection weight can be adjusted adaptively with the test sample. Five vehicles with different loads are considered in the actual engineering test when the vehicle speed is ranged from 10 km/h to 50 km/h while the temperature shifts from 16℃ to 29℃.According to 500 cycle tests,the RBF neural network model of vehicle weigh-in-motion is constructed.The experiment results show that the averaged weighing er-ror of the proposed algorithm is less than 0.06% and the averaged value of the maximum real-time is 0.022 3,meet-ing effectively the practical engineering requirements.
展开▼
