基于SNN-LSTM的小样本数据下轴承故障诊断方法

Bearing fault diagnosis method based on SNN-LSTM under limited samples

基于深度学习的故障诊断方法的实现,需要用到大量的、有标注的训练样本,而在小样本数据下,采用这些方法会产生模型欠拟合问题,同时获得的分类准确率也较低。为了解决上述问题,提出了一种小样本数据下结合孪生神经网络(SNN)与长短时记忆网络(LSTM)的轴承故障诊断方法。首先,以一对带有正负标签的原始振动信号样本作为诊断方法的输入,采用比较二者相似度的方法,扩充了训练样本个数;然后,采用共享提取样本对特征网络参数的方法,完成了SNN的搭建过程;使用卷积层、池化层及LSTM层提取了原始振动信号的特征,通过计算二者之间的曼哈顿距离,判断输入样本对的相似度,对不同状态下的轴承完成了分类;最后,为了验证基于SNN-LSTM的故障诊断方法在轴承故障诊断中的有效性,通过轴承故障诊断实验,采集了在不同转速、不同状态下的轴承振动信号数据。研究结果表明:当样本数量仅为140个,采用基于SNN-LSTM的故障诊断方法的准确率达到80.57%,相比于深度学习经典方法,在小样本数据下采用该方法具有更高的诊断准确率。

The realization of fault diagnosis methods based on deep learning required the use of a large number of labeled training samples,and in the case of small sample data,the use of these methods would cause the problem of model underfitting,and the classification accuracy obtained was also low.In order to solve the above problems,a bearing fault diagnosis method combining Siamese neural network(SNN)and long and short time memory network(LSTM)was proposed under limited samples.Firstly,a pair of original vibration signal sample pairs with positive or negative labels were used as the input of the diagnosis method,and the number of training samples could be expanded by comparing the similarity between a pair of input samples.Then,the method of sharing the network parameters of extracting the features of sample pairs was used and the construction process of SNN was completed.The convolution layer,pooling layer and the LSTM layer were used to extract the features of the original vibration signal.The similarity of a pair of input samples was judged by Manhattan distance between them and the bearings under different conditions were classified.Finally,in order to verify the effectiveness of the fault diagnosis method based on SNN-LSTM in bearing fault diagnosis,a test bench for bearing fault diagnosis was completed,and the bearing vibration signals under different speeds and different states were collected.The research results indicate that the accuracy of the proposed method reaches 80.57%when the number of samples is only 140,which is higher than the classical deep learning method under limited samples.