压缩感知和改进深层小波网络在轴承故障诊断中的应用

APPLICATION OF COMPRESSIVE SENSING AND IMPROVED DEEP WAVELET NEURAL NETWORK IN BEARING FAULT DIAGNOSIS

针对传统滚动轴承故障诊断方法过度依赖专家经验和故障特征提取困难的问题,提出一种基于压缩感知(Compressive Sensing,CS)和改进深层小波神经网络(Deep Wavelet Neural Network,DWNN)方法。首先对采集到的轴承振动信号进行CS降噪并压缩采样;其次设计改进小波自编码器(Wavelet Auto-Encoder,WAE)进而构造DWNN,并引入"跨层"连接缓解网络的梯度消失现象;最后利用大量无标签轴承压缩数据对DWNN进行无监督预训练并利用少量带标签数据对网络有监督微调,进而实现故障判别。实验结果表明提出方法能够有效地对轴承进行多种故障类型和多种故障程度的识别,受先验知识和主观影响较小,避免了复杂的人工特征提取过程,特征提取能力和识别能力优于人工神经网络、深度信念网络、深度稀疏自编码器等模型。

Aiming at the problems of traditional bearings fault diagnosis methods had such shortcomings as largely dependent on expert prior knowledge and difficulty in fault feature extraction,a method based on compressive sensing(CS)and improved deep wavelet neural network(DWNN)was proposed.Firstly,the collected vibration data of bearings were de-noised and compressed by CS.Secondly,the improved wavelet auto-encoder was designed to construct the DWNN,and the"crosslayer"connection was introduced to alleviate the gradient disappearance of the network.Finally,unsupervised pre-training of DWNN was performed using a large amount of unlabeled compressed data and supervised and fine-tuned with a small amount of tagged data to realize fault discrimination.The experimental results show that the method can effectively identify the bearings with multiple fault types and multiple fault severities,which is less affected by prior knowledge and subjective knowledge and avoids complex artificial feature extraction process.The feature extraction ability and recognition ability of proposed method are superior than artificial neural network,deep belief network,deep sparse auto-encoder and so on.