基于MTF-MSMCNN的小样本滚动轴承故障诊断方法

A rolling bearing fault diagnosis method based on MTF-MSMCNN with small sample

针对样本数量不足以及工况条件复杂导致故障识别精度低下的问题,提出一种基于马尔科夫转移场与多维监督卷积神经网络(Markov transition field and multidimensional supervised module convolutional neural networks,MTF-MSMCNN)的小样本滚动轴承故障诊断方法。采用MTF编码方式将一维滚动轴承信号转化为二维特征图像,使其保留时间相关性;提出多维监督模块(Multidimensional supervision module,MSM),在空间维度和通道维度监测重要故障特征并自适应赋予权重,提升模型捕捉关键特征的能力;将MSM嵌入到卷积神经网络中,构建出MSMCNN模型;通过试验构建复杂工况条件,将MTF图像输入到所提模型进行故障诊断,并运用两种数据集验证模型有效性。试验结果表明,MTF-MSMCNN在每类故障训练集样本仅有10个且在0 dB噪声污染下故障诊断精度依然可达90%左右,对比其他诊断模型,本文所提方法在小样本、变工况以及噪声干扰条件下具有更高的识别准确率、更强的泛化能力以及抗噪性能。

Considering the problem of low fault identification accuracy caused by insufficient sample size and complex working conditions,a fault diagnosis method of rolling bearing with small sample based on Markov transition field and multidimensional supervised module convolutional neural network(MTF-MSMCNN)was proposed.The one-dimensional rolling bearing signal was transformed into two-dimensional feature image using MTF coding method to preserve temporal correlation.Multidimensional supervision module(MSM)was presented to monitor important fault features in both spatial and channel dimension and assign weight adaptively,which can improve the model’s ability to capture key features.MSM was embedded into the convolutional neural network to build a MSMCNN model.The complex working conditions were constructed through experiments,and the MTF images were input into the MTF-MSMCNN network model for fault diagnosis.Two data sets were used to verify the model validity.The experimental results showed that the MTF-MSMCNN had only 10 samples in each type of fault training set,and its fault diagnosis accuracy can still reach about 90%under 0 dB noise pollution.Compared with other diagnostic models,the method proposed had higher recognition accuracy,stronger generalization ability and anti-noise performance under the conditions of small samples,variable working conditions and noise interference.