Abstract: To solve the problems of low diagnosis accuracy and poor generalization performance caused by different sample distribution and insufficient samples,a fault diagnosis method for rolling bearings under small samples and variable working conditions based on MTF and SSCAM-MSCNN was proposed.The Markov Transition Field (MTF) was used to transform one-dimensional vibration signal into two-dimensional feature map with time-dependent.A novel Stripe Self-calibrating Attention Mechanism (SSCAM) was put forward,which could not only enhance feature extraction ability in long-distance direction,but also establish inter-channel dependence and capture global effective information.Then,SSCAM was introduced into Multi-Scale Convolutional Neural Network (MSCNN) to construct SSCAM-MSCNN model.Finally,the MTF two-dimensional feature map was input into the proposed model for training,and the optimized network model was used to test and output the classification results.The proposed method was validated by both Case Western Reserve University data set and MFS rolling bearing data set of our laboratory.Meanwhile,the MFS data was noised up and compared with other fault diagnosis models.The experimental results showed that the proposed method had higher recognition accuracy,stronger generalization performance and anti-noise performance under small samples and variable operating conditions.

Key words: rolling bearing, Markov transition field, convolutional neural network, stripe self-calibrating attention mechanism, small sample, fault diagnosis

摘要： 针对滚动轴承在不同工况条件下样本分布不同以及故障样本数量不足导致故障诊断精度低、泛化性能差的问题,提出一种小样本下基于MTF与SSCAM-MSCNN的滚动轴承变工况故障诊断方法。首先,运用马尔科夫转移场(MTF)将一维振动信号转化为具有时间相关性的二维特征图。其次,提出条纹自校正注意力机制(SSCAM),它不仅可以加强模型在长距离方向上的特征提取能力,还能建立通道间依赖关系,可以对全局有效信息进行捕捉。然后,将SSCAM引入到多尺度神经网络(MSCNN)中,构建出SSCAM-MSCNN模型。最后,将MTF二维特征图输入到所提模型中进行训练,采用优化后的网络模型进行测试并输出分类结果。通过美国凯斯西储大学以及本实验室MFS滚动轴承数据集对所提方法进行验证,同时对后者进行加噪处理,与其他故障诊断模型进行对比。试验结果表明,所提方法在小样本、变工况条件下具有更高的识别精度、更强的泛化性能与抗噪性能。

关键词: 滚动轴承, 马尔科夫转移场, 卷积神经网络, 条纹自校正注意力机制, 小样本, 故障诊断