摘要
针对传统预测模型处理复杂非线性变形数据能力不足的问题,应用一种深度学习模型——长短期记忆网络(Long Short-Term Memory,LSTM)进行变形预测分析.研究发现LSTM隐藏层受有限变形数据序列长度的影响较大,针对此类问题,文中在LSTM的基础上提出了一种改进模型.改进的模型将具有多个隐藏层的神经网络放在LSTM模型前,采用全连接层将两种网络连接起来,旨在提高预测精度.改进的LSTM模型主要对时间序列数据进行预测,以西南部矿业某矿区自动监测数据为例,将LSTM模型和改进后的LSTM模型进行了对比分析实验.为验证多变量输入是否会对预测精度造成影响,构建基于改进LSTM的多点预测模型,并与改进LSTM的单点预测对比.结果表明,LSTM可用于变形预测,且改进的LSTM模型很好地改善了LSTM存在的问题,预测精度相对更高,适用于多点预测,可批量处理变形数据.
As the traditional prediction model cannot efficiently deal with complicated nonlinear deformation data, this paper presents a kind of deef learning model—Long Short-Term Memory(LSTM)to predict and analyse deformation. It showed the hidden layer of LSTM is greatly affected by the length of the finite deformation data sequence. To solve such a problem, this paper proposes a better LSTM model on the basis of LSTM. The improved model puts multiple hidden layers in front of the LSTM model and connects these two kinds of neural network with fully connected layer to improve prediction accuracy. The improved LSTM model is mainly used to predict time series data. This paper compares the LSTM model with the improved one on dealing with the automatic monitoring data of a mining area in the southwestern China. To verify whether multivariate inputs have an impact on prediction accuracy, the paper makes a contrast between multipoints and single point prediction of improved LSTM. The results show that LSTM can be used for deformation prediction,and the improved one can well improve the existing problems of LSTM. Its prediction accuracy is relatively high, which is suitable for multi-point prediction and batch processing of deformation data.
引文
[1]陈林,黄腾,郑浩.传统小波神经网络在变形预测中的模型改进[J].测绘科学,2017,42(9):1-7.
[2]仲洁.基于神经网络的地铁结构安全监测与分析[D].南京:东南大学,2016.
[3]周永胜,姚殿梅.多元评价体系组合模型在铁路隧道变形预测中的应用[J].隧道建设,2017,37(6):676-683.
[4] Lecun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553):436-444.
[5]王鑫,吴际,刘超,等.基于LSTM循环神经网络的故障时间序列预测[J].北京航空航天大学学报,2018,44(4):772-784.
[6] LI X, Peng L, YAO X J, et al. Long short-term memory neural network for air pollutant concentration predictions:Method development and evaluation[J]. Environmental Pollution, 2017, 231(1):997-1004.
[7]陈亮,王震,王刚.深度学习框架下LSTM网络在短期电力负荷预测中的应用[J].电力信息与通信技术,2017,15(5):8-11.
[8] Zhang Q, Wang H, Dong J, et al. Prediction of sea surface temperature using long short-term memory[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(10):1745-1749.
[9]朱乔木,李弘毅,王子琪,等.基于长短期记忆网络的风电场发电功率超短期预测[J].电网技术,2017,41(12):3797-3802.
[10] Cho K, Van Merri觕nboer B, Gulcehre C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[J]. Computer Science, 2014,arxiv:1406.1078.
[11]保罗,郭旭琦,乔铁柱,等.改进LSTM神经网络在磨机负荷参数软测量中的应用[J].中国矿山工程,2017,46(3):66-69.
[12]张冲.基于Attention-Based LSTM模型的文本分类技术的研究[D].南京:南京大学,2016.
[13] Graves A, Mohamed A R, Hinton G. Speech recognition with deep recurrent neural networks[C]//IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2013:6645-6649.
[14] Sutskever I. Training recurrent neural networks[D]. Toronto:University of Toronto, 2013.
[15]王国栋.基于LSTM的舰船运动姿态短期预测及仿真研究[D].镇江:江苏科技大学,2017.
[16]吴兵兵.基于词向量和LSTM的汉语零指代消解研究[D].哈尔滨:哈尔滨工业大学,2016.
[17] Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Computation, 1997, 9(8):1735-1780.
[18]孙瑞奇.基于LSTM神经网络的美股股指价格趋势预测模型的研究[D].北京:首都经济贸易大学,2016.
[19] Dalton Barnaby, Sozubek Serdar, Saldana Manuel, et al. Reduction of parameters in fully connected layers of neural networks[P]:US2017337463, 2017-11-23.
[20]袁秋壮,魏松杰,罗娜.基于深度学习神经网络的SAR星上目标识别系统研究[J].上海航天,2017,34(5):46-53.
[21] Williams R J, Peng J. An efficient gradient-based algorithm for on-line training of recurrent network trajectories[J]. Neural Computation,1990,2(4):490-501.
[22] Kinga D, Adam J B. A method for stochastic optimization[C]//International Conference on Learning Representations(ICLR),2015.