基于Resnet网络和Attention-GRU门控机制的滚动轴承故障诊断

针对轴承故障诊断中大多现有方法特征提取复杂且诊断方法不是端到端等问题,结合深度学习理论,提出了一种基于Resnet网络(残差网络)和Attention机制(注意力机制)的轴承故障诊断方法。诊断思想是:首先,通过Resnet网络对输入的滚动轴承的一维振动时序信号进行特征提取;其次,将特征提取后的特征图经过Map-to-sequence操作将特征图转换为特征序列送入到Attention机制的GRU(门控循环单元)网络中进行预测;最后,通过分类器将预测后的结果分类输出即可得到诊断结果。实验表明,该模型对各故障类别的诊断率均在98%以上,模型诊断准确率普遍优于其他传统的诊断方法,相较于一些最近流行的基于深度学习轴承故障诊断方法效果也提升显著。

Rolling Bearing Fault Diagnosis Method Based on Resnet Network and Attention-GRU Gating Mechanism

For the fault diagnosis of most existing methods in bearing fault diagnosis, and the diagnosis method is not end-to-end, we combine a deep learning theory to propose a bearing fault diagnosis method based on Resnet network(residual network) and Attention mechanism. The diagnosis idea is as follows: Firstly, the one-dimensional vibration timing signal of the input rolling bearing is extracted by Resnet network;secondly, the feature map after feature extraction is transformed into the feature sequence by the Map-to-sequence operation. The GRU(Gated Cycle Unit) neural network of the mechanism performs prediction;finally, the predicted result is classified and output by the classifier to obtain a diagnosis result. Experiments show that the diagnostic rate of the model is more than 98% for each fault category. The accuracy of model diagnosis is generally better than other traditional diagnostic methods. Compared with some recent popular deep learning bearing fault diagnosis methods, the effect is also significantly improved.