基于特征扩展CapsNet的轴承故障诊断模型

Bearing Fault Diagnosis Model Based on Feature Extended CapsNet

中文关键词:  噪声抑制  特征扩展  混合约束  余弦相似度度量  CapsNet  故障诊断

英文关键词:noise suppression  feature expansion  hybrid constraint  cosine similarity measure  CapsNet  fault diagnosis

中文摘要:

      基于深度网络的轴承故障诊断模型，其深层次的特征提取往往需要大量的训练样本，诊断模型应用于小样本故障信号时受到限制。利用胶囊网络(CapsNet)在处理小样本数据上的优势，通过扩展故障特征向量和改进预测胶囊迭代约束条件，使网络适用于噪声工况下的轴承故障诊断，提高预测精度。在CapsNet的初级胶囊层引入特征图跨通道约束关系，由原来的通道内约束改进为通道内-通道间混合约束，扩展胶囊特征向量，使其能描述故障信号时间序列中与远距离序列点相关的几何特性，为数字胶囊层预测网络提供更完备的故障特征。同时，利用余弦相似度作为特征向量的度量并为数字胶囊层的迭代筛选提供依据，避免向量模值造成的分类误差。凯斯西储大学(CWRU)轴承数据集实验结果表明，改进的 iCapsNet 诊断模型泛化性能得到明显提升，在信噪比为 0 dB 时，预测精度可达到 90.9%，相比原 CapsNet 模型提高了44.8%。

英文摘要:

      The extraction of the deep features for the bearing diagnosis model based on the neural networks relies greatly on the huge training samples.The diagnostic performance of the model would suffer an application limitation because of the small number of the fault samples.The capsule network(CapsNet) is adopted for the construction of the bearing diagnosis model due to superiority in dealing with the small number of the samples.With the feature vector expansion of the fault and constraint improvement of the capsules in the CapsNet model,the prediction accuracy of the diagnosis model can be guaranteed when the bearing fault diagnosis task under the noise condition is involved.A cross-channeled constraint relation is introduced to the primary capsule layer of the CapsNet.Then the CapsNet diagnosis model can describe the geometric properties of the time sequence fault signal which are dependent on the long-distance points.Consequently,more complete fault features are available for the following prediction digital capsule network by the substitution of the original inner-channeled constraint with a hybrid constraint.Moreover,in order to avoid the fault classification error caused by the modulus confusion of the feature vector,the cosine similarity measure is adopted as the iterative basis for the digital capsule layer.The experimental results on Case Western Reserve University dataset (CWRU) show that the generalization performance of the improved iCapsNet diagnostic model has been significantly improved.Under 0 dB noise interference,the prediction accuracy can reach 90.9%,which is 44.8% higher than the original CapsNet model.