噪声干扰下的CCSD-CNN轴承故障诊断方法

Fault Diagnosis Method for Bearings Based on CCSD-CNN Under Noise Interference

针对噪声干扰下卷积神经网络(CNN)出现性能退化、故障识别准确率降低的问题,利用循环相关熵谱密度(CCSD)能有效抑制高斯和非高斯噪声的特点,提出了一种基于CCSD-CNN的滚动轴承故障诊断方法。首先,将一维振动信号转变为二维循环相关熵谱密度,优化卷积神经网络输入信号质量;然后,将循环相关熵谱密度作为卷积神经网络的输入层,通过二维卷积神经网络实现轴承故障特征提取和故障分类;最后,利用凯斯西储大学轴承数据集进行验证,并与基于短时傅里叶变换(STFT)、连续小波变换(CWT)和谱相关密度(SCD)的信号预处理方法进行对比。试验结果表明:CCSD-CNN方法提高了噪声环境下卷积神经网络的鲁棒性和诊断准确率,性能优于STFT-CNN,CWT-CNN和SCD-CNN。

Aimed at the problems of performance degradation and fault recognition accuracy reduction of convolutional neural network(CNN)under noise interference,a fault diagnosis method for rolling bearings based on CCSD-CNN is proposed by using cyclic correntropy spectral density(CCSD),which can effectively suppress Gaussian and non-Gaussian noise.Firstly,the one-dimensional vibration signal is transformed into two-dimensional CCSD,and the quality of input signal of CNN is optimized.Then,the CCSD is used as input layer of CNN,the fault feature extraction and fault classification of the bearings are realized by two-dimensional CNN.Finally,the bearing dataset from Case Western Reserve University is used for verification,and compared with signal preprocessing methods based on short-time Fourier transform(STFT),continuous wavelet transform(CWT)and spectral correlation density(SCD).The experimental results show that CCSD-CNN method improves the robustness and diagnostic accuracy of CNN under noisy environments,its performance is better than that of STFT-CNN,CWT-CNN and SCD-CNN.