首先,建立卷积神经网络、深度置信网络、支持向量机和以一维卷积神经网络全连接层特征为输入的支持向量机模型(1DCNN-SVM),对比上述模型在地铁车轮失圆状态分类识别上的效果;其次,利用代理模型构建轴箱垂向加速度均方根与车速和多边形磨耗幅值之间的映射关系;最后,通过智能优化算法逆向求解幅值,对比不同代理模型和智能优化算法在多边形磨耗幅值识别上的适用性.研究结果表明:1DCNN-SVM模型在正常、低阶多边形、高阶多边形、随机非圆车轮4类典型的车轮不圆度状态分类识别中取得99.82%的准确性,相比另外3种分类方法,其泛化性能和强化学习能力都具有明显的优势.在车轮多边形磨耗幅值识别方面,基于克里金模型(KSM)和粒子群算法(PSO)的波深识别模型具有更好的预测稳定性和时效性.
Firstly,convolutional neural network,deep belief network,support vector machine and support vector machine model with the full connection layer features of a one-dimensional convolutional neural network as input(1DCNN-SVM)were established respectively.Secondly,the effects of the above models on the classification of out-of-roundness of metro wheels were compared.The mapping relationship between the root mean square of the vertical acceleration of the axle box and the vehicle speed and the polygonal wear amplitude was constructed by surrogate models.Finally,the wheel polygonal wear amplitude was inversely solved by the intelligent optimization algorithm.The applicability of different surrogate models and intelligent optimization algorithms was compared in the identification of the wheel polydonal wear amplitude.The results show that the 1DCNN-SVM model achieves a classification rate of 99.82%in four types of typical wheel out-of-roundness,such as normal,low-order polygons,high-order polygons and non-periodic non-roundness wheels.Compared with the other three classification methods,its generalization performance and reinforcement learning ability have obvious advantages.In terms of wheel polygonal wear amplitude identification,the method based on Kriging model(KSM)and particle swarm optimization algorithm(PSO)has better prediction stability and timeliness.