针对同一识别算法下不同损伤类型以及损伤程度识别准确率低的问题，提出一种多尺度融合特征与一维卷积神经网络(One-dimensional Convolution Neural Network，1D-CNN)相结合的车轮损伤诊断方法。利用完全噪声辅助聚合经验模态分解(Complete EEMD with Adaptive Noise，CEEMDAN)和魏格纳-维尔分布(Wigner-Ville Distribu-tion，WVD)联合时频分析方法，对轴箱加速度信号进行融合特征提取，将提取到的多维度融合特征作为样本输入，构建适合车轮踏面损伤诊断的CNN模型，对样本中的不同损伤类型和损伤程度进行分类识别。经仿真分析和实验验证表明:所提出的多维度融合特征对于不同车速下的损伤类型以及损伤程度都有很好的识别能力，识别准确率可达到98％，且鲁棒性强，可为车轮踏面损伤识别和评估提供新的方法。
Wheel Defect Recognition Method Based on Fused Features and CNN
To solve the problem of low recognition accuracy of different damage types and damage degrees in the same model,a wheel tread damage diagnosis method based on multi-scale characteristics and 1D-CNN was proposed.Using the joint time-frequency analysis method of complete EEMD with adaptive noise (CEEMDAN) and Wigner-Ville distribution (WVD),the fusion feature extraction was performed on the axle box acceleration signal.Then,the extracted multi-dimen-sional fusion features were used as sample inputs to construct a CNN model suitable for wheel tread damage diagnosis,and different types and degrees of damage in the samples were classified and identified.Through simulation analysis and experi-mental verification,it is shown that the proposed multi-dimensional features have good recognition ability for damage types and degrees at different vehicle speeds,with a recognition accuracy of 98％ and strong robustness.This work provides a new method for identifying and evaluating wheel tread damage.
万方数据
维普
