Abstract: Aiming at the problems of low accuracy and poor generalization under the condition of small samples due to the excessive dependence of traditional deep learning models on a large amount of training data, a small sample bearing fault diagnosis method based on improved siamese neural network is proposed. This method is dedicated to using the metric learning to judge the similarity between samples. Compared with complex and multi-layer deep models, its structure is simpler and it is easier to train with small samples. Specifically, this paper designs three modules to realize fault diagnosis. In the feature extraction stage, the long short-term memory network and the convolutional neural network are used to extract the temporal and spatial features of the fault signal, which can make full use of the limited sample information; the relationship measurement stage replaces the common Euclidean distance measurement with an adaptive network measurement method and uses global mean pooling to reduce network parameters; in addition, compared with the standard siamese network, the constructed fault classification network can directly identify the faults of the samples, without tedious sample comparison in the testing stage. The experimental results show that the ISNN method has better fault diagnosis accuracy and generalization performance than various comparative methods under the limited number of training samples.

Key words: rolling bearing, siamese neural network, small sample, fault diagnosis

摘要： 针对传统深度学习模型过度依赖大量训练数据，在小样本条件下会出现准确率低和泛化性差的问题，提出一种基于改进孪生神经网络（improved siamese neural network，ISNN）的小样本轴承故障诊断方法。该方法致力于利用度量学习思想来判断样本间的相似性，相比于复杂、多层的深度模型其结构更简单，在小样本下也更容易训练。具体而言，共设计了三个模块来实现故障诊断。特征提取阶段分别利用长短时记忆网络和卷积神经网络提取出故障信号的时间和空间特征，能够充分利用有限的样本信息；关系度量阶段以自适应的网络度量方式来代替常用的欧式距离度量，并且引入全局均值池化来减少网络参数；此外，与标准孪生网络相比，构建的故障分类网络可以直接对样本进行故障识别，在测试阶段无须繁琐的样本比对。实验结果表明，ISNN方法在有限的训练样本数量下，故障诊断准确率和泛化性能均优于各种对比方法。

关键词: 滚动轴承, 孪生网络, 小样本, 故障诊断