【作者】 李旭；

【导师】 何巍华（David He）；

【作者基本信息】 东北大学 ， 机械设计及理论， 2019， 硕士

【摘要】 随着数据科学的发展,有越来越多的研究者开始尝试使用深度学习技术进行寿命预测。剩余寿命预测任务的关键是准确地预测设备的退化过程,然后根据退化趋势推算得到剩余寿命。因此,历史寿命数据的分析就就变得非常重要,只有从历史数据准确地分析性能退化的过程,才能实现更加准确的剩余寿命预测。而深度学习作为一种数据驱动技术,其不需要精确的数学模型,并且非常适合用来做大数据分析。本文正是基于此背景,基于卷积自动编码器和卷积神经网络进行剩余寿命的预测的研究,主要研究内容包括:(1)本文综述了基于深度学习技术的剩余寿命预测方法,着重介绍了卷积神经网络和循环神经网络在剩余寿命预测领域的应用,详细阐述了基于深度学习技术的寿命预测模型常用的损失函数及评价指标。(2)提出了基于卷积自动编码器和卷积神经网络的剩余寿命预测算法。该模型通过卷积自动编码器对输入信号进行预特征提取,然后使用卷积网络对信号做进一步的特征提取,并回归特征与输出之间的关系。本文通过添加损伤比这个特征,简化了网络的学习过程,提高了模型的准确度。(3)设计并实施了中碳钢板材的拉伸疲劳试验。本文的疲劳试验以中碳钢板材为试件,通过拉伸疲劳试验机施加远小于屈服极限强度的载荷来模拟材料在低应力水平下的高周疲劳过程。同时,编写了基于LAbVIEW的数据采集系统来进行数据采集。(4)本文以中碳钢的疲劳试验中采集到的声发射信号数据为数据集,对本文的提出的模型进行了性能评估,并且针对本文中采用的一些技术对性能的影响进行了分析。实验结果表明,本文提出的剩余寿命预测算法具有较好的准确度。通过基于卷积自动编码器和卷积神经网络的剩余寿命预测算法对中碳钢板材剩余寿命的预测,研究了传统疲劳累积理论与深度网络的结合,为更加准确的剩余寿命预测算法的研究提供了一种思路。更多还原

【Abstract】 With the development of data science,more and more researchers have begun to try to use deep learning technology for life prediction.The key to the remaining life prediction task is to accurately predict the degradation process of the equipment,and then derive the remaining life based on the degradation trend.Therefore,the analysis of historical life data becomes very important.Only the process of accurately analyzing performance degradation from historical data can achieve more accurate residual life prediction.Deep learning,as a data-driven technology,does not require an accurate mathematical model and is well suited for big data analysis.Based on this background,this paper studies the prediction of remaining life based on convolution autoencoder and convolutional neural network.The main research contents include:(1)This paper reviews the residual life prediction method based on deep learning technology,and focuses on the application of convolutional neural network and cyclic neural network in the field of residual life prediction.The loss function and evaluation index of life prediction model based on deep learning technology are elaborated in detail.(2)A residual life prediction algorithm based on convolution autoencoder and convolutional neural network is proposed.The model pre-features the input signal through a convolutional auto-encoder,and then uses the convolution network to further feature extraction of the signal and regress the relationship between the feature and the output.By adding the damage ratio feature,this paper simplifies the learning process of the network and improves the accuracy of the model.(3)The tensile fatigue test of medium carbon steel sheets was designed and implemented.In the fatigue test of this paper,a medium carbon steel plate is used as a test piece,and a load far less than the yield limit strength is applied by a tensile fatigue tester to simulate a high cycle fatigue process of the material at a low stress level.At the same time,a data acquisition system based on LAbVIEW was written for data acquisition.(4)In this paper,the acoustic emission signal data collected in the fatigue test of medium carbon steel is taken as the data set.The performance of the proposed model is evaluated.The effects of some techniques used in this paper on performance are analyzed.The experimental results show that the residual life prediction algorithm proposed in this paper has better accuracy.By predicting the remaining life of medium carbon steel sheets based on the residual life prediction algorithm based on convolution autoencoder and convolutional neural network,the combination of traditional fatigue accumulation theory and depth network is studied,which provides a more accurate study of residual life prediction algorithm.A way of thinking.更多还原