基于Ce-ATRP-CNN的滚动轴承故障诊断方法

A rolling bearing fault diagnosis method based on Ce-ATRP-CNN

摘要:

针对滚动轴承故障振动信号在实际运行环境下故障特征易被噪声影响,一维信号转换为二维图像时故障冲击性纹理特征易被噪声纹理特征所掩盖的问题,创新性地将自适应阈值融入递归图编码过程,提出了一种名为自适应阈值递归图编码(Adaptive Threshold Recurrence Plot, ATRP)的图像转换方法；在此基础上,进一步构建了Ce - ATRP - CNN故障诊断模型。通过倒谱分析(Cepstrum,Ce)对原始振动信号预处理,可以一定程度上滤除噪声；应用提出的ATRP将处理后的一维振动信号转换为二维故障特征图像—ATRP图；将ATRP图作为卷积神经网络(Convolutional Neural Network, CNN)的输入实现对滚动轴承故障特征的深入提取和识别诊断；应用CWRU和HUST轴承故障数据集进行了试验验证,故障诊断准确率分别达到了99.17%和99.05%；实验结果表明该模型具有良好的故障识别能力和泛化性。

Abstract:

Aiming at the problem that the fault features of vibration signals from rolling bearing were easily affected by noise in the actual operating environment, and the texture features of fault impact were easily covered by texture features of noise when converting one-dimensional signals into two-dimensional images, an image conversion method called adaptive threshold recurrence plot (ATRP) was proposed by introducing the adaptive threshold into the recurrence plot encoding. On this basis, a Ce-ATRP-CNN fault diagnosis model was further constructed. The original vibration signals were pre-processed through cepstrum analysis, which could filter out noise to a certain extent. The one-dimensional vibration signals after processing were converted into two-dimensional fault feature images, namely ATRP images, using the ATRP method proposed in this paper. The ATRP images were used as the input of the convolutional neural network (CNN) to deeply extract and identify the fault features of rolling bearings for diagnosis. The model proposed was verified through experiments using the bearing fault datasets from CWRU and HUST. The fault diagnosis accuracies for the two datasets reached 99.17% and 99.05% respectively. The experimental results indicate that the model has good fault recognition ability and generalization performance.