基于分数阶傅里叶域特征融合的多工况管道堵塞 识别方法

Method for Identifying Pipeline Blockage Based on Fractional FourierDomain Feature Fusion under Multiple Operating Conditions

针对排水管道声学检测的实际应用中,声学信号在特征选择缺乏指导的情况下容易提取过高信息重叠度的高维特征导致不同工况管道堵塞识别率低下的问题,基于分数阶傅里叶变换(Fractional Fourier Transform,FRFT)原理,提出一种基于FRFT和多核学习(Multiple Kernel Learning,MKL)特征融合的管道多重堵塞识别方法。该方法首先把难以区分的原始数据映射到多阶次分数阶傅里叶变换时频平面,然后计算各阶次的样本熵作为区分度量特征。运用MKL自动学习FRFT样本熵特征的系数,将分数域中阶次选择问题转换为多核网络中的系数交替优化问题,挖掘特征的深层含义,最终将这些信息进行多特征融合,实现了多工况管道堵塞识别。实验结果表明:在与不同阶次集合的融合特征对比后,最终的融合特征提高了不同类别间样本的区分度,能够有效识别复杂运行状态下多工况管道中的堵塞物、三通件和管道尾端,识别准确率达到95%,在多工况管道堵塞识别中相较于传统特征融合模型识别率显著提高。

In practical applications of acoustic detection for drainage pipelines,lack of guidance in feature selection makes the extraction of high-dimensional features with high information overlap and it may cause low recognition rates in identifying different blockages in pipelines under different operating conditions.Through basing on Fractional Fourier Transform(FRFT)and Multiple Kernel Learning(MKL)feature fusion,a pipeline multi-blockage identification method was proposed.In which,having the indistinguishable raw data mapped to the time-frequency plane of multiple orders of fractional Fourier transform,and then having the sample entropy of each order calculated as a discriminative feature measure.Through using MKL to automatically learn the coefficients of the FRFT sample entropy features,the order selection in the fractional domain was transformed into an alternating optimization problem in the multiple kernel network,including having the deep meaning of features explored and the multi-feature fusion ultimately achieved to identify multi-condition pipeline blockage.Experimental results show that,compared with the fusion features from different order sets,the final fusion feature can improve the discrimination between different categories of samples,and can effectively identify blockages,T-joints,and pipe ends in complex operating conditions within multi-condition pipelines;and the recognition accuracy can reache 95%,which is a significant improvement compared to traditional feature fusion models in multi-condition pipeline blockage identification.