一种融合特征与卷积神经网络的车轮缺陷识别方法  

Wheel Defect Recognition Method Based on Fused Features and CNN

作　　者：尹兆珂 缪炳荣[1] 张盈[1] 袁哲锋 胡天棋 YIN Zhaoke;MIAO Bingrong;ZHANG Ying;YUAN Zhefeng;HU Tianqi(State Key Laboratory of Rail Transit Vehicle System,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学轨道交通运载系统全国重点实验室,成都610031

出　　处：《噪声与振动控制》2024年第6期149-155,共7页Noise and Vibration Control

基　　金：四川省重点研发资助项目(2023YFG0197);牵引动力国家重点实验室自主研发资助项目(2023TPL_T08);中央高校基本科研业务费-专题研究资助项目(2682022ZTPY007)。

摘　　要：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 Distribution,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。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-dimensional 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 experimental 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.

关 键 词：故障诊断 轴箱加速度 CEEMDAN-WVD联合时频分析法 一维卷积神经网络 

分 类 号：U211.5[交通运输工程—道路与铁道工程] TP181[自动化与计算机技术—控制理论与控制工程]