基于YOLOX的分布式PCCP断丝自动监测方法研究

Research on YOLOX-based Distributed PCCP Automatic Wire Breakage Monitoring Method

为了研究一种针对预应力钢筒混凝土管(PCCP)断丝信号的识别技术,设计了使用DN4000mm的PCCP建成模拟测试环境,在非常接近实际工况的情况下进行了1∶1原型断丝监测试验。试验将采集到的断丝信号与之前采集到的运行管道内的噪声信号相互组合,分别通过连续小波变换(CWT)和同步挤压小波变换(SWT)两种方式转化的时频谱图作为断丝信号数据集,采用基于YOLOX的目标检测算法通过提取数据集中断丝信号时频谱图图像特征来判断断丝事件的发生。两组数据集训练出模型的准确率、召回率、F1分数,误检率均达到100%、100%、1和0%。随后的剪枝计算阶段,在保证检测精度不变的前提下,SWT组的剪枝率高于CWT组,最终模型大小仅为1.36 MByte。可见,通过SWT变换得到的时频谱图可以在不损失精度的前提下更大的简化YOLOX模型,为开发PCCP断丝监测系统的嵌入式部署提供了参考。

In order to study a wire breakage identification technique for prestressed steel cylinder concrete pipe(PCCP),this paper designs a simulated test environment by using a DN 4000mm PCCP and conducts a 1∶1 sample wire breakage monitoring test under almost actual working conditions.The test combines the collected wire break signal and the previously collected noise signal in the operating pipeline with each other,and the time-frequency spectrograms transformed by both continuous wavelet transform(CWT)and synchrosqueezed wavelet transform(SWT)are used as the wire break signal dataset.The YOLOX-based target detection algorithm is used to determine the occurrence of wire breakage events by extracting the features of the time-frequency spectrograms of the wire breakage signals in the dataset.The accuracy rate,recall rate,F1 score,and false detection rate of the model are 100%,100%,1 and 0%for both datasets trained.In the subsequent pruning calculation stage,the pruning rate of the SWT group is higher than that of the CWT group,while ensuring the same detection accuracy,and the final model size is only 1.36 MByte.It can be seen that the time-frequency spectrogram obtained by SWT transformation can simplify the YOLOX model without loss of accuracy,which provides a reference for the development of embedded installation of PCCP broken wire monitoring system.