基于深度残差收缩网络的滚动轴承故障诊断

滚动轴承的准确故障诊断是确保机械设备安全可靠运行的必要手段。针对多故障、长时间序列的滚动轴承振动信号，提出了一种基于深度残差收缩网络(DRSN)模型的故障诊断方法。首先，根据采集到的滚动轴承数据构造故障样本，针对多种故障类型下的长时间序列的振动信号，按照一定尺寸将长时间序列矩阵化，构成多故障类型的灰度图故障样本。从正常到故障的滚动轴承性能退化过程，通过多个采样点的随机采样，构造全寿命周期的故障样本用于故障诊断。其次，在多层深度学习模型基础上，将残差收缩网络模块加入到卷积神经网络(CNN)中构建深度残差收缩网络模型用于故障诊断，其中通过将残差项加入到网络中训练解决了多层网络模型的模型退化问题，利用软阈值化实现了样本降噪。最后，为了验证所提方法的有效性，采集了滚动轴承的多故障时间序列样本和全寿命周期故障样本用于故障诊断。实例验证的结果表明:所提深度残差收缩网络模型在处理含噪声样本时仍具有良好的鲁棒性，多层网络模型下没有明显的网络退化，能够保持较高的故障诊断正确率。在处理2种轴承故障数据集时，与其他模型相比，所提方法训练误差更低，平均故障诊断正确率提高1%~6%。 

Fault diagnosis of rolling bearing based on deep residual shrinkage network

Accurate fault diagnosis of rolling bearing is a necessary means to ensure the safe and reliable operation of mechanical equipment. In this paper, a fault diagnosis method based on Deep Residual Shrinkage Network (DRSN) is proposed for the vibration signal of rolling bearing with multiple faults and long time series. Firstly, fault samples are constructed according to the collected rolling bearing data. For the vibration signals of long time series under various fault types, the long time series are matrixed according to a certain size, so as to form the gray image fault samples of multiple fault types. Aimed at the performance degradation process of rolling bearings from normal to fault, the whole life cycle fault samples are constructed for fault diagnosis through random sampling of multiple sampling points. Secondly, based on the multi-layer deep learning model, the residual shrinkage network module is added to the Convolutional Neural Network (CNN) to build the deep residual shrinkage network model, in which the model degradation problem of the multi-layer network model is solved by adding the residual term to the network training, and the sample noise reduction is realized by using soft thresholding. Finally, in order to verify the effectiveness of the proposed method, multi-fault time series samples and life cycle fault samples of rolling bearing are collected for fault diagnosis. The result of the example shows that the proposed model has good robustness under the noise interference, there is no obvious network degradation under the multi-layer network model, and it can maintain a high accuracy of fault diagnosis. When dealing with two kinds of bearing fault datasets, compared with other models, this method has lower training error, and the average accuracy of fault classification is increased by 1%-6%.