基于改进的LSTM的药品温湿度预测方法
|
摘要
针对医药冷链系统中药品温湿度数据不易诊断的问题,提出一种改进的长短期记忆(LSTM)预测药品温湿度的方法。首先通过插值扩充算法扩充湿度数据集,接着提出一种内含多个LSTM细胞元的LSTM结构,代替传统的迭代预测,随后通过Adam优化算法调整网络参数和改变网络层数降低预测误差,实现对药品温湿度的提前预判。最后在药店冷藏柜中采集到的药品温湿度数据集上进行测试,均方误差(MSE)为0. 036 9。与传统的BP神经网络预测方法和高斯过程混合模型预测方法对比,改进的LSTM药品温湿度预测方法预测更准确。
Aiming at the problem that the temperature and humidity data of medicines in medical cold chain system are not easy to diagnose,an improved long short-term memory( LSTM) method for predicting the temperature and humidity of drugs is proposed. The method first expands the humidity data set by interpolation expansion algorithm,and then proposes an LSTM structure containing multiple LSTM cell elements instead of the traditional iterative prediction. Then the Adam optimization algorithm adjusts the network parameters and changes the network layer to reduce the prediction error. Achieve early prediction of the temperature and humidity of the drug.Finally,the test was carried out on the temperature and humidity data set of the drug collected in the pharmacy refrigerator. The mean square error was 0. 036 9. Compared with the traditional BP neural network prediction method and the Gaussian process mixture model prediction method,the improved LSTM drug temperature and humidity prediction method is more accurate.
Aiming at the problem that the temperature and humidity data of medicines in medical cold chain system are not easy to diagnose,an improved long short-term memory( LSTM) method for predicting the temperature and humidity of drugs is proposed. The method first expands the humidity data set by interpolation expansion algorithm,and then proposes an LSTM structure containing multiple LSTM cell elements instead of the traditional iterative prediction. Then the Adam optimization algorithm adjusts the network parameters and changes the network layer to reduce the prediction error. Achieve early prediction of the temperature and humidity of the drug.Finally,the test was carried out on the temperature and humidity data set of the drug collected in the pharmacy refrigerator. The mean square error was 0. 036 9. Compared with the traditional BP neural network prediction method and the Gaussian process mixture model prediction method,the improved LSTM drug temperature and humidity prediction method is more accurate.
引文
[1]黄奕斌.我国医药冷链物流发展现状及建议[J].物流技术与应用,2017,22(S1):35-37.HUANG Y B. The status and suggestions of thedevelopment of pharmaceutical cold chain logistics inChina[J]. Logistics Technology and Application,2017,22(S1):35-37.
[2]龚小龙,孔玲爽,袁川来,等.非线性时间序列粒子群优化B样条网络预测模型[J].电子测量与仪器学报,2017,31(12):1890-1895.GONG X L,KONG L SH,YUAN CH L,et al. Nonlineartime series particle swarm optimization B-spline networkprediction model[J]. Journal of Electronic Measurementand Instrumentation,2017,31(12):1890-1895.
[3]朱江淼,宋文峰,高源,等.基于改进型BP神经网络的氢原子钟钟差预测[J].仪器仪表学报,2016,37(2):454-460.ZHU J M,SONG W F,GAO Y,et al. Prediction ofhydrogen atom clock difference based on improved BPneural network[J]. Chinese Journal of ScientificInstrument,2016,37(02):454-460
[4]张一凡,余小清,安炫东.基于改进SVM的纳税评估和预测[J].电子测量技术,2016,39(8):79-84.ZHANG Y F,YU X Q,AN X D. Tax evaluation andprediction based on improved SVM[J]. ElectronicMeasurement Technology,2016,39(8):79-84.
[5]葛珂楠,程忠庆,赵云鹏.基于BP神经网络的地下空间温湿度预测[J].计算机工程与设计,2013,34(6):2190-2194.GE K N,CHENG ZH Q,ZHAO Y P. Temperature andhumidity prediction of underground space based on BPneural network[J]. Computer Engineering and Design,2013,34(6):2190-2194.
[6]周亚同,赵翔宇,何峰,等.基于高斯过程混合模型的大气温湿度预测[J].农业工程学院,2018,34(5):219-226.ZHOU Y T,ZHAO X Y,HE F,et al. Prediction of largetemperature and humidity based on Gaussian processhybrid model[J]. Transactions of the Chinese Society ofAgricultural Engineering,2018,34(5):219-226.
[7] LECUN Y,BENGIO Y,HINTON G. Deep learning[J].Nature,2015,521(7553):436-436.
[8] GRAVES A. Long short-term memory[C]. SupervisedSequence Labelling with Recurrent Neural Networks,Springer Berlin Heidelberg,2012:1735-1780.
[9] DONAHUE J,HENDRICKS L A,ROHRBACH M,etal. Long-term recurrent convolutional networks for visualrecognition and description[J]. IEEE Transactions onPattern Analysis&Machine Intelligence,2017,39(4):677-691.
[10] VINYALS O,TOSHEV A,BENGIO S,et al. Show andtell:A neural image caption generator[C]. IEEEConference on Computer Vision and Pattern Recognition,IEEE Computer Society,2015:3156-3164.
[11] MA X,TAO Z,WANG Y,et al. Long short-termmemory neural network for traffic speed prediction usingremote microwave sensor data[J]. TransportationResearch Part C,2015(54):187-197.
[12]徐尧强,方乐恒,赵冬华,等.基于LSTM神经网络的用电量预测[J].贵州电力技术,2017,20(8):25-29.XU Y Q,FANF Y H,ZHAO D H,et al. Electricityconsumption forecast based on LSTM neural network[J].Guizhou Electric Power Technology, 2017, 20(8):25-29.
[13]王鑫,吴际,刘超.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.WANG X,WU J,LIU CH. Fault time series predictionbased on LSTM recurrent neural network[J]. Journal ofBeijing University of Aeronautics and Astronautics,2018,44(4):772-784.
[14]李泽龙,杨春节,刘文辉.基于LSTM-RNN模型的铁水硅含量预测[J].化工学报,2018(3):992-997.LI Z L,YANG CH J,LIU W H. Prediction of siliconcontent in molten iron based on LSTM-RNN model[J].CIESC Journal,2018(3):992-997.
[15] HU Y,HUBER A,ANUMULA J,et al. Overcoming thevanishing gradient problem in plain recurrent networks[C].ARXIV,2018.
[16] GRAVES A,SCHMIDHUBER J. Framewise phonemeclassification with bidirectional LSTM and other neuralnetwork architectures[J]. Neural Networks, 2005,18(5):602-610.
[17] YEUNG S,RUSSAKOVSKY O,NING J,et al. Everymoment counts:Dense detailed labeling of actions incomplex videos[J]. International Journal of ComputerVision,2018,126(2-4):375-389.
[2]龚小龙,孔玲爽,袁川来,等.非线性时间序列粒子群优化B样条网络预测模型[J].电子测量与仪器学报,2017,31(12):1890-1895.GONG X L,KONG L SH,YUAN CH L,et al. Nonlineartime series particle swarm optimization B-spline networkprediction model[J]. Journal of Electronic Measurementand Instrumentation,2017,31(12):1890-1895.
[3]朱江淼,宋文峰,高源,等.基于改进型BP神经网络的氢原子钟钟差预测[J].仪器仪表学报,2016,37(2):454-460.ZHU J M,SONG W F,GAO Y,et al. Prediction ofhydrogen atom clock difference based on improved BPneural network[J]. Chinese Journal of ScientificInstrument,2016,37(02):454-460
[4]张一凡,余小清,安炫东.基于改进SVM的纳税评估和预测[J].电子测量技术,2016,39(8):79-84.ZHANG Y F,YU X Q,AN X D. Tax evaluation andprediction based on improved SVM[J]. ElectronicMeasurement Technology,2016,39(8):79-84.
[5]葛珂楠,程忠庆,赵云鹏.基于BP神经网络的地下空间温湿度预测[J].计算机工程与设计,2013,34(6):2190-2194.GE K N,CHENG ZH Q,ZHAO Y P. Temperature andhumidity prediction of underground space based on BPneural network[J]. Computer Engineering and Design,2013,34(6):2190-2194.
[6]周亚同,赵翔宇,何峰,等.基于高斯过程混合模型的大气温湿度预测[J].农业工程学院,2018,34(5):219-226.ZHOU Y T,ZHAO X Y,HE F,et al. Prediction of largetemperature and humidity based on Gaussian processhybrid model[J]. Transactions of the Chinese Society ofAgricultural Engineering,2018,34(5):219-226.
[7] LECUN Y,BENGIO Y,HINTON G. Deep learning[J].Nature,2015,521(7553):436-436.
[8] GRAVES A. Long short-term memory[C]. SupervisedSequence Labelling with Recurrent Neural Networks,Springer Berlin Heidelberg,2012:1735-1780.
[9] DONAHUE J,HENDRICKS L A,ROHRBACH M,etal. Long-term recurrent convolutional networks for visualrecognition and description[J]. IEEE Transactions onPattern Analysis&Machine Intelligence,2017,39(4):677-691.
[10] VINYALS O,TOSHEV A,BENGIO S,et al. Show andtell:A neural image caption generator[C]. IEEEConference on Computer Vision and Pattern Recognition,IEEE Computer Society,2015:3156-3164.
[11] MA X,TAO Z,WANG Y,et al. Long short-termmemory neural network for traffic speed prediction usingremote microwave sensor data[J]. TransportationResearch Part C,2015(54):187-197.
[12]徐尧强,方乐恒,赵冬华,等.基于LSTM神经网络的用电量预测[J].贵州电力技术,2017,20(8):25-29.XU Y Q,FANF Y H,ZHAO D H,et al. Electricityconsumption forecast based on LSTM neural network[J].Guizhou Electric Power Technology, 2017, 20(8):25-29.
[13]王鑫,吴际,刘超.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.WANG X,WU J,LIU CH. Fault time series predictionbased on LSTM recurrent neural network[J]. Journal ofBeijing University of Aeronautics and Astronautics,2018,44(4):772-784.
[14]李泽龙,杨春节,刘文辉.基于LSTM-RNN模型的铁水硅含量预测[J].化工学报,2018(3):992-997.LI Z L,YANG CH J,LIU W H. Prediction of siliconcontent in molten iron based on LSTM-RNN model[J].CIESC Journal,2018(3):992-997.
[15] HU Y,HUBER A,ANUMULA J,et al. Overcoming thevanishing gradient problem in plain recurrent networks[C].ARXIV,2018.
[16] GRAVES A,SCHMIDHUBER J. Framewise phonemeclassification with bidirectional LSTM and other neuralnetwork architectures[J]. Neural Networks, 2005,18(5):602-610.
[17] YEUNG S,RUSSAKOVSKY O,NING J,et al. Everymoment counts:Dense detailed labeling of actions incomplex videos[J]. International Journal of ComputerVision,2018,126(2-4):375-389.