小样本下基于并行异构网络的变工况纸机轴承故障诊断方法

Fault Diagnosis Method of Paper Machine Bearings for Variable Working Conditions Based on Parallel Heterogeneous Network in Small Samples

中文关键词:  并行异构CNN  纸机轴承  轴承故障诊断

Key Words:parallel heterogeneous CNN  paper machine bearings  bearing fault diagnosis

中文摘要:

      传统纸机轴承故障诊断模型在实际应用中，存在由于故障振动信号数据量小、信号有效信息占比低导致的变工况下故障诊断准确度下降等问题。针对此问题，本课题提出一种小样本下基于并行异构网络的变工况纸机轴承故障诊断方法。首先，将源域和目标域信号分别转换为相应的格拉姆角场矩阵、马尔科夫变迁场矩阵和欧氏距离矩阵，并对所得的3种矩阵逐行进行交叉组合，以此作为网络输入；其次，基于卷积神经网络（CNN）对2D-CNN进行改进，设计融合注意力机制的多通道并行异构网络，实现对信号深层特征的自动提取；然后，基于对抗思想设计域判别器与分类器，通过多核最大均值差异（MK-MMD）对两域特征边缘分布进行对齐，实现对变工况下轴承故障的识别。最后，分别基于凯斯西储大学滚动轴承数据集与实验室自建纸机轴承故障模拟平台采集数据，进行迁移学习实验验证。结果表明，该纸机轴承故障迁移学习网络模型具有优异的特征挖掘能力，对变工况下的纸机轴承故障具有较高的识别精度。

Abstract:

      In practical applications， the traditional paper machine bearing fault diagnosis model has problems， such as decreased accuracy in fault diagnosis under variable working conditions due to the small amount of fault vibration signal data and low proportion of effective signal information. In response to this issue， this project proposed a fault diagnosis method of paper machine bearings for variable working conditions based on parallel heterogeneous network in small samples. Firstly， the source and target domain signals were converted into corresponding Gram angle field matrix， Markov transition field matrix， and Euclidean distance matrix， respectively. The obtained three matrices were cross combined row by row and used as network inputs. Secondly， based on convolutional neural network （CNN）， 2D-CNN was improved by designing a multi-channel parallel heterogeneous network that integrated attention mechanism to automatically extract deep features of signals. Then， based on adversarial thinking， domain discriminators and classifiers were designed to align the feature edge distributions of the two domains through multi kernel maximum mean difference （MK-MMD）， achieving recognition of bearing faults under variable operating conditions. Finally， transfer learning experiments were conducted to verify the data collected from the Case Western Reserve University rolling bearing dataset， and the laboratory self-built paper machine bearing fault simulation platform. The results indicated that the paper machine bearing fault transfer learning network model had excellent feature mining ability and high recognition accuracy for paper machine bearing faults under variable operating conditions.