相对于传统的“对信号进行特征提取+人工选择对数据敏感的特征值+预测模型”的滚动轴承寿命预测方法，深度信念网络（Deep Belief Networks，DBN）有显著的优势：DBN可以直接处理原始数据，让机器自动学习信号特征，从而免去了特征提取和选择的过程，提高了预测的智能性。但是传统的DBN采用固定学习率进行网络学习，不利于寻找最优结果。基于此，本文提出了一种改进的深度信念网络——全参数动态学习深度信念网络（Global parameters dynamic learning Deep Belief Networks，GPDLDBN），并将其应用于滚动轴承寿命预测中。GPDLDBN预测模型由多层受限玻尔兹曼机（Restricted Boltzmann Machines，RBM）单元组成，首先采用自下而上的逐层无监督贪婪算法训练参数；然后采用自上而下的监督学习算法微调整个网络参数，两个过程均采用新的全参数动态学习策略设置各参数。采用GPDLDBN预测模型对实测的滚动轴承寿命数据进行了预测，并与传统的固定学习率的DBN预测模型进行了对比分析。结果表明，GPDLDBN预测模型能够有效加快收敛速度，减少模型的训练时间，且具有更高的预测精度。
Abstract
Compared with the traditional methods with the mode of "Extract the features of signal+Select the features that are sensitive to data by artificial+Prediction model" for the rolling bearing life prediction, the deep belief networks (DBN) have significant advantages.The DBN handles raw data directly, and the machine automatically learns the signal features, instead of extracting features and selecting features, therefore, it improves the intelligence of life prediction.However, the traditional DBN adopts the fixed learning rate for network learning, which is not conducive to finding the optimal result.Targeting this disadvantage, an improved deep belief network, by the name of global parameters dynamic learning deep belief networks(GPDLDBN), was proposed, and applied to the prediction of rolling bearing life.The GPDLDBN prediction model consists of a multi-layer Restricted Boltzmann Machines unit.First, parameters of a greedy algorithm were trained layer-by-layer without supervision from the bottom to up.Then, a top-down supervised learning algorithm was used to fine-tune the entire network parameters.The global dynamic learning strategy was adopted for the parameters setting up in both processes.The total life data of rolling bearing were predicted based on the GPDLDBN prediction model, and compared with the DBN prediction model with traditional fixed learning rate.The results show that the GPDLDBN prediction model can effectively improve the convergence speed,reduce the training time and has higher prediction accuracy.
关键词
深度学习 /
全参数动态学习深度信念网络 /
滚动轴承 /
寿命预测
{{custom_keyword}} /
Key words
Deep learning;Global parameters dynamic learning deep belief networks;Rolling bearing /
Life prediction
{{custom_keyword}} /
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] 张小丽，王保建，马 猛等. 滚动轴承寿命预测综述[J]. 机械设计与制造. 2015, (10): 221-224.
Zhang Xiao-li, Wang Bao-jian, Ma Meng. Rolling bearing life prediction overview[J]. Mechanical Design and Manufacturing.2015, (10): 221-224.
[2] 李巍华，单外平，曾雪琼. 基于深度信念网络的轴承故障分类识别[J]. 振动工程学报. 2016, 29(2): 340-347.
Li Wei-hua, Shan Wai-ping, Zeng Xue-qiong. Classification and identification of bearing fault based on deep belief networks[J]. Iournal of Vibration Engineering. 2016, 29(2): 340-347.
[3] Tian Yang. The Application of Improved Deep Belief Network in Surface Roughness of Grinding[J]. Modular Machine Tool&Automatic Manufacturing Technique. 2016(7): 108-110.
[4] 尹宝才，王文通，王立春. 深度学习研究综述[J].北京工业大学学报. 2015, 41(1): 48-59.
Yi Bao-cai, Wang Wen-tong, Wang li-chun. Review of deep learning[J]. Journal of Beijing Technology University. 2015, 41(1): 48-59.
[5] Liang Chen, Jun-chi Zhang, Na Wang. Deep belief networks based popularity prediction for online video servicess[J/OL]. http://www.cnki.net/kcms/detail/11.2127.TP.20160120.1503.010html, 2016.
[6] Xing Wang, Yi-peng Zhou, Dong-qing Zhou. Research on Low Probability of Intercept Radar Signal Recognition Using Deep Belief Network and Bispectra Diagonal Slice[J/OL]. http://www.cnki.net/kcms/detail/11.4494.TN.20161010.1358.034.html, 2016.
[7] Li-min Zhang, Kai Liu. Document Features Extraction Based on DBM[J]. MICROELECERONICS&COMPUTER, 2015, 32(2): 142-147.
[8] Salakhutdinov R, Murray I. On the quantitative analysis of deep belief networks[C]//Proceedings of the 25th international conference on Machine learning. ACM, 2008: 872-879.
[9] Fischer A, Igel C. An introduction to restricted Boltzmann machines[M]//Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications. Springer Berlin Heidel berg, 2012: 14-36.
[10] Tran V T, Al Thobiani F, Ball A. An approach to fault diagnosis of reciprocating compressor valves using Teager-Kaiser energy operator and deep belief networks[J]. Expert Systems with Applications, 2014, 41(9): 4113-4122.
[11] Y Bengio. Learning Deep Architectures for AI[Z]. Foundations&Trends in Machine Learning, 2009．
[12] 田杨. 改进的深度信念网在磨削加工粗糙度值预测中的应用[J]. 组合机床与自动化加工技术. 2016, (7): 108-110.
Tian Yang. The application of improved deep belief network in the prediction of grinding roughness value[J]. Combined machine tool and automation machining technology. 2016, (7): 108-110.
[13] Hai Qiu, Jay Lee, Jing Lin. Wavelet Filter-based Weak Signature Detection Method and its Application on Roller Bearing Prognostics. Journal of Sound and Vibration 289 (2006) 1066-1090
{{custom_fnGroup.title_cn}}
脚注
{{custom_fn.content}}
PDF(1872 KB)
