基于集成学习的铁路隧道空洞敲击检查声音识别  

Percussion inspection voice recognition of railway tunnel voids based on ensemble methods

作　　者：高磊 刘振奎[1] 张昊宇 魏晓悦 张奎 GAO Lei;LIU Zhenkui;ZHANG Haoyu;WEI Xiaoyue;ZHANG Kui(School of Civil Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学土木工程学院,兰州730070

出　　处：《振动与冲击》2022年第14期58-63,83,共7页Journal of Vibration and Shock

基　　金：国家自然科学基金(11662007,51268031);兰州市科技计划项目(2018-4-33)。

摘　　要：隧道衬砌空洞敲击检查方法是目前铁路隧道中应用最多的检查方法,但其空洞识别和数据的记录均靠人工完成。为实现铁路隧道空洞敲击检查声音智能识别,将采集到的隧道敲击检查音频文件进行预处理,选取645个声音样本,提取24维梅尔频率倒谱系数(Mel frequency cepstrum coefficients,MFCC)作为声音样本的声学特征参数,通过集成算法(梯度提升决策树GBDT)训练样本声学特征,建立隧道空洞敲击检查声音分类模型,最后将该模型应用于实际铁路隧道空洞敲击检查声音识别分类。实例研究表明:与优化的支持向量机(cross-validation-support vector machine,CV-SVM)模型和改进径向基神经网络(particle swarm optimization algorithm-radial basis function neural network,PSO-RBF)模型相比,GBDT集成算法模型具有更高的准确率和更少的运算时间,在面对异常数据时具有更强的稳定性,能够准确地根据铁路隧道空洞敲击检查声音诊断衬砌后是否存在空洞。The percussion inspection method for voids behind tunnel lining is widely used in railway tunnel at present,but the identification and data recording are still completed manually.In order to recognize the percussion inspection voice in railway tunnel intelligently,the voice files which were collected from the tunnel percussion inspection were preprocessed,645 groups of voice samples were selected,and the 24-dimensional Mel frequency cepstrum coefficients(MFCC)were extracted as the acoustic feature parameters of the voice samples.A classification model for the voice of tunnel percussion inspection was established by training the parameters with the ensemble algorithm(gradient boosting decision tree,GBDT).Finally,the model was applied to the actual railway tunnel to recognise voids according to the percussion inspection voice.The case study shows that the GBDT ensemble methods model has higher accuracy and less operation time,compared with the optimized support vector machine(CV-SVM)model and the improved radial basis function neural network(PSO-RBF)model.It has stronger stability when processing abnormal data,and can accurately diagnose the existence of voids behind the lining according to the percussion inspection voice.

关 键 词：铁路隧道 声音识别 梅尔频率倒谱系数(MFCC) 梯度提升决策树 支持向量机(SVM) 改进RBF神经网络 

分 类 号：U25[交通运输工程—道路与铁道工程]