基于深度学习的矿井滚动轴承故障诊断方法

A fault diagnosis method for mine rolling bearings based on deep learning

  • 摘要: 针对传统卷积神经网络在煤矿井下等复杂环境中难以充分挖掘数据特征等问题,提出了一种基于马尔可夫转移场(MTF)和双通道多尺度卷积胶囊网络(DMCCN)的矿井滚动轴承故障诊断方法,构建了MTF−DMCCN故障诊断模型。根据MTF和灰度图对原始振动信号进行编码后,采用双通道输入模式连接卷积网络获取浅层特征;将特征图进行融合后输入到胶囊网络,提高模型对空间信息的敏感度;在网络中引入Inception模块,聚焦多尺度特征,加强网络的特征提取能力;通过胶囊层进行向量化处理,实现滚动轴承的故障诊断与分类。消融实验、抗噪性及泛化性实验结果表明:Inception模块、灰度图输入、MTF图像输入均对轴承故障诊断具有正向促进的作用,MTF编码对模型的诊断精度提升最高;MTF−DMCCN模型具有较好的鲁棒性和抗噪声能力;MTF−DMCCN模型具有优异的变转速适应能力,在不同工况条件下具有良好的泛化性能。为进一步验证模型性能,选取格拉姆角差场(GADF)、格拉姆角和场(GASF)、灰度图、MTF等图像编码方式与不同网络相结合,采用辛辛那提大学数据集(IMS)进行对比实验,结果表明,MTF−DMCCN模型能有效识别滚动轴承故障类型,平均故障诊断准确率达99.37%。

     

    Abstract: A fault diagnosis method for mine rolling bearings based on Markov transition field(MTF) and dual-channel multi-scale convolutional capsule network (DMCCN) is proposed to address the problem of traditional convolutional neural networks being unable to fully explore data features in complex environments such as coal mines. The MTF-DMCCN fault diagnosis model is constructed. After encoding the original vibration signal based on MTF and grayscale image, a dual channel input mode is used to connect the convolutional network to obtain shallow features. The method inputs the feature maps fusion into the capsule network to improve the sensitivity of the model to spatial information. The method introduces Inception modules into the network to focus on multi-scale features and enhance the network's feature extraction capabilities. Finally, vectorization processing is carried out through the capsule layer to achieve fault diagnosis and classification of rolling bearings. The results of ablation, noise resistance, and generalization experiments show that the Inception module, grayscale image input, and MTF image input all have a positive promoting effect on bearing fault diagnosis. MTF coding has the highest improvement in diagnostic precision of the model. The MTF-DMCCN model has good robustness and noise resistance. The MTF-DMCCN model has excellent adaptability to variable speed and still exhibits good generalization performance under different operating conditions. To further validate the performance of the model, image encoding methods such as Gram angle difference field (GADF), Gram angle sum field (GASF), grayscale image, and MTF are selected and combined with different networks. Comparative experiments are conducted using the University of Cincinnati intelligent maintenance system (IMS). The results show that the MTF-DMCCN model can effectively recognize the type of rolling bearing faults, with an average fault diagnosis accuracy of 99.37%.

     

/

返回文章
返回