基于无监督深度学习的声发射信号聚类分析

为提高声发射信号检测诊断的自动化程度,直接从声发射波形出发,提出了一种基于深度神经网络与聚类分析的声发射信号分类方法。针对声发射信号标签数据难以获取的问题,采用无监督学习方式,根据大量声发射实测波形进行深度一维卷积自编码器训练,实现了声发射信号特征的自动提取,进而结合K均值聚类算法准确区分不同类型的声发射信号。铅芯在复合材料板上突然断裂和摩擦的声发射试验表明,提出的方法能自动识别不同类别的声发射信号,识别效果优于基于人为设定声发射信号特征的聚类方法。

Clustering analysis of acoustic emission signals based on unsupervised deep learning

Based on acoustic emission(AE)waveform,this paper proposes a method of AE signal classification based on deep neural network and cluster analysis.In order to avoid the difficulty of making AE signal label manually,a large number of measured AE waveforms are used for deep one-dimensional convolutional autoencoder training in an unsupervised learning method to realize the automatic extraction of AE signal characteristics,and then the k-means clustering algorithm is combined to accurately distinguish different types of AE signals.The AE experiments of sudden fracture and friction of lead core on composite plates show that the proposed method can automatically identify different types of AE signals,and the identification accuracy is higher than the clustering method based on the artificial characteristics of AE signals.