滚动轴承是机械设备中的重要零件,其工作状态直接关系着设备的运行,一旦发生故障会引起整个设备的正常运行,甚至引发重大的安全事故,因此,对其剩余寿命预测对设备的健康管理具有重要意义.提出了一种基于自编码-长短时记忆网络(autoencoder-long short term memory,AELSTM)迁移学习(transfer learning,TL)的滚动轴承剩余寿命预测方法,首先采用自动编码器自动提取源域中原始振动信号中的特征,再构建双层LSTM模型对剩余寿命进行预测,通过源域中训练获得AELSTM模型,再用目标域中的数据对AELSTM模型训练,完成对模型参数的微调,最后用调整好的模型对目标域中的数据进行预测.通过参数共享和微调两种方法,大大简化了模型在目标域上的训练过程.试验结果表明,在同轴承不同工况下,所提出模型相比于其他4种迁移学习方法的均方根误差分别降低了45.9%、58.9%、42.8%以及83.8%;在不同轴承不同工况下,所提出模型的均方根误差分别降低了16.9%、18.9%、11.7%以及8.9%.
Rolling bearing remaining life prediction based on AELSTM and model transfer learning
Rolling bearings are important components in mechanical equipment,and their operational status is directly related to the operation of the equipment.When a failure occurs,it can disrupt the normal operation of the entire device,potentially leading to significant safety accidents.Therefore,predicting the remaining lifespan of these bearings is of crucial significance for equipment health management.This paper introduces a method for predicting the remaining lifespan of rolling bearings based on the transfer of an AELSTM model.Firstly,an autoencoder is utilized to automatically extract features from the raw vibration signals in the source domain.Subsequently,a two-layer LSTM model is constructed to predict the remaining lifespan.The AELSTM model is trained in the source domain and then fine-tuned with data from the target domain to adjust the model parameters.Finally,the adjusted model is used to predict the data in the target domain.By employing parameter sharing and fine-tuning methods,the training process of the model in the target domain is greatly simplified.The experimental results indicate that,under different operating conditions of the same bearing,the proposed model exhibited a reduction in root mean square error compared to four other transfer learning methods by 45.9%,58.9%,42.8%,and 83.8%respectively.In the case of different bearings under various operating conditions,the proposed model demonstrated reductions by 16.9%,18.9%,11.7%,and 8.9%respectively.
remaining useful life predictionmodel transferlong short term memoryautoencoder
NETL
NSTL
万方数据
