摘要
针对轴承故障数据样本少、未知故障难以分类等问题,提出了一种将一维卷积神经网络(1D convolutional neural network, 1D-CNN)连接深层长短时记忆循环神经网络(Deep long-short-term memory neural network, DLSTM)的模型结合迁移学习的故障诊断方法。该诊断方法基于电机振动数据,利用CNN提取故障特征;将特征作为DLSTM的输入,进一步学习、编码从CNN中学习的特征序列信息,捕获高级特征用于故障分类;首先用充足的西储轴承数据对该故障诊断模型进行预训练,再利用迁移学习放松训练数据和测试数据可不必独立同分布的能力,使用自制实验平台的小样本数据微调预训练模型。最后用迁移学习后的模型,对跨工况、跨型号、跨故障的故障轴承数据进行模拟实验。结果表明,所提出的方法与其他方法相比鲁棒性强,训练速度更快,能够更精确的诊断故障,平均诊断精度达到99%以上。
A bearing fault diagnosis method that connects a 1D convolutional neural network(1D-CNN)to a model of deep long short-term memory recurrent neural network(DLSTM)combined with transfer learning is proposed to address the problems of small fault data samples and difficulty in classifying unknown faults.The diagnostic method is based on motor vibration data and uses CNN to extract fault features;the features are used as input to the DLSTM,which further learns and encodes the feature sequence information learned from the CNN to capture high-level features for fault classification;the fault diagnosis model is first pre-trained with sufficient samples of Case Western Reserve University data,and then uses transfer learning to relax the training data and test data that do not need to be independently and identically distributed.The pre-training model was then fine-tuned using small samples of data from a home-made experimental platform.Finally,the transfer learning model is used to simulate experiments on faulty bearing data across operating conditions,models and faults.The results show that the proposed method is more robust and faster to train than other methods,and can diagnose faults more accurately,with an average diagnostic accuracy of over 99%.
作者
仇芝
徐泽瑜
陈涛
石明江
韦明辉
QIU Zhi;XU Zeyu;CHEN Tao;SHI Mingjiang;WEI Minghui(School of Electrical and Mechanical Engineering,Southwest Petroleum University,Chengdu 610500,China)
出处
《机械科学与技术》
北大核心
2025年第2期288-297,共10页
Mechanical Science and Technology for Aerospace Engineering
基金
国家自然科学基金项目(51804267)
四川省科技厅计划项目(2019YJ0318)。
关键词
小样本数据集故障诊断
卷积神经网络
长短期记忆网络
迁移学习
small sample data set fault diagnosis
convolutional neural network
long short-term memory network
transfer learning
