基于声发射时频分析与卷积神经网络的液膜密封摩擦状态识别

Friction State Recognition of Liquid Film Seal Based on Acoustic Emission Time-frequency Analysis and Convolution Neural Network

针对液膜密封状态监测领域无损监测开发不足、信号特征评估困难以及摩擦状态判别智能化特性缺乏的问题,提出一种基于声发射时频分析与卷积神经网络的液膜密封摩擦状态识别方法。该方法将声发射无损监测技术应用于液膜密封的摩擦状态监测,卷积神经网络作为液膜密封摩擦状态自主决策的实现手段,声发射信号的时频信息作为卷积神经网络的特征输入,分析短时傅立叶变换、S变换以及小波变换3种时频分析方法对卷积神经网络识别性能的影响。结果表明:对于液膜密封的声发射信号,3种时频分析方法与卷积神经网络结合的优选顺序为:短时傅立叶变换、S变换、小波变换;基于声发射时频分析与卷积神经网络的液膜密封摩擦状态识别方法准确率较高,相比其他识别方法取得了较好的识别效果。

Aimed at the problems of insufficient development of nondestructive testing,difficult in evaluation of signal characteristics and lack of intelligent characteristics of friction state discrimination in the field of liquid film seal condition monitoring,a friction state recognition method of liquid film seal based on acoustic emission time-frequency analysis and convolution neural network(CNN)was proposed.In this method,acoustic emission nondestructive testing technology was applied to the friction state monitoring of liquid film seals,convolutional neural network was used as the means of autonomous decision-making of friction state of liquid film seals,and the time-frequency information of acoustic emission signals was used as the feature input of convolution neural network.The influence of three time-frequency analysis methods of short-time Fourier transform,S transform and wavelet transform on the recognition performance of convolution neural network was investigated.The results show that for the acoustic emission signal of liquid film seal,the optimal order of the combination of three time-frequency analysis methods and convolutional neural network is short-time Fourier transform,S transform,wavelet transform.The recognition method of friction state recognition of liquid film seal based on acoustic emission time-frequency analysis and convolution neural network has high accuracy,and can achieve good recognition results compared with other recognition methods.