【作者】 张珊珊；

【导师】 林连雷；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2020， 硕士

【摘要】 随着武器装备试验环境的多样化与虚拟化,为提高对武器装备进行虚拟试验的真实性和准确度,在虚拟试验中添加虚拟大气环境是必要的。如何生成虚拟大气环境数据成为了一个关键研究课题。虚拟大气环境数据生成本质是一个时间序列问题,近年来,深度学习在处理大量时序数据方面有着优越的表现。本文利用深度学习生成虚拟大气环境数据,所做具体研究如下。首先,为了得到深度学习技术所用的训练数据集、测试数据集,以及提高网络模型准确度等,对大气环境数据预处理方法进行研究。研究处理大气环境数据中的离群值和缺失值的方法,确保深度学习模型生成的结果不会因为这两者的影响出现偏差。研究大气环境数据标准化方法,以加快网络训练速度。最后研究大气环境数据时间滑窗处理方法,对德国耶拿数据集进行处理得到深度学习模型所需的训练集,验证集和测试集。其次,为了快速准确地生成大气环境数据,基于大气环境中多种参数具有物理相关性的特点,不同于大多数传统的单参数方法,本文提出合并的长短时记忆网络(M-LSTM)模型。利用具有较强相关性的多个大气参数生成特定的大气参数,提高准确度。文中以温度为例,利用数据相关性分析方法分析大气环境中其他参数与温度之间的关系,将基于不同参数生成温度的多个LSTM网络合并,充分利用原始数据包含的特征。最终实验结果表明,M-LSTM-AH模型在测试集上的均方误差至少比经典LSTM-AH模型降低9.3%,实现了对温度数据的准确生成。最后,针对联合试验环境需要多参数的大气环境数据的问题,本文提出了LSTM-GRU-multi网络模型。由于大气环境数据样本具有特征多,数据量巨大的特点,单独使用LSTM模型训练需要大量时间。GRU与LSTM相比结构更简单,网络参数减少进而收敛速度更快。本文将LSTM和GRU相结合,在LSTM有效处理大量大气环境数据的前提下,加快网络训练速度。同时由于单层的网络结构过于简单,远远不能满足处理大气环境数据多特征的需求,本文采用多层的网络结构。实验结果表明,LSTM-GRU-multi模型比传统多层LSTM模型在测试集上的MSE至少降低了11.65%,同时在训练时间上也具有2.3%的优势,LSTM-GRU-multi模型能准确且快速生成多参数大气环境数据。更多还原

【Abstract】 With the diversification and virtualization of weapon equipment test environments,in order to improve the authenticity and accuracy of virtual tests on weapon equipment,it is necessary to add a virtual atmosphere environment to the virtual test.How to generate virtual atmospheric environment data has become an important research topic.The generation of virtual atmospheric environment data is essentially a time series problem.In recent years,deep learning has performed well in processing large amounts of time series data.This article uses deep learning to generate virtual atmospheric environment data.The specific research done is as follows.First,in order to obtain the training data set and test data set used by the deep learning technology,and to improve the accuracy of the network model,etc.,the atmospheric environment data preprocessing method is studied.The method of dealing with outliers and missing values in atmospheric environment data is studied to ensure that the results generated by the deep learning model will not be biased due to the influence of the two.The standardization method of atmospheric environment data is studied to speed up the network training speed.Finally,the time sliding window processing method of atmospheric environment data is studied,and the training set,verification set and test set required by the deep learning model are obtained by processing the Jena data set in Germany.Secondly,in order to quickly and accurately generate atmospheric environmental data,based on the characteristics of multiple parameters in the atmospheric environment with physical correlation,unlike most traditional singleparameter methods,this paper proposes a merged long-short-term memory network(M-LSTM)model.Use multiple atmospheric parameters with strong correlation to generate specific atmospheric parameters to improve accuracy.Taking temperature as an example in this paper,the relationship between other parameters in the atmospheric environment and temperature is analyzed using data correlation analysis methods,and multiple LSTM networks that generate temperatures based on different parameters are combined to make full use of the features contained in the original data.The final experimental results show that the mean square error of the test set of the M-LSTM-AH model is at least 9.3% lower than that of the classic LSTM-AH model and accurate temperature data generation is achieved.Finally,for the problem that the multi-parameter atmospheric environment data is needed in the joint test environment,this paper proposes the LSTM-GRU-multi network model.Because the atmospheric environment data samples have many characteristics and huge data volume,it takes a lot of time to train using the LSTM model alone.Compared with LSTM,GRU has a simpler structure,and the network parameters are reduced to converge faster.This article combines LSTM and GRU to speed up the network training speed under the premise that LSTM effectively processes a large amount of atmospheric environmental data.At the same time,because the single-layer network structure is too simple,it is far from meeting the needs of processing multiple characteristics of atmospheric environmental data.This paper uses a multi-layer network structure.Multiple sets of experiments on the Jena data set show that the LSTM-GRU-multi model reduces the MSE on the test set by at least 11.65% compared with the traditional multi-layer LSTM model,and also has a 2.3% advantage in training time.LSTM-GRU-multi model can accurately and quickly generate multi-parameter atmospheric environment data.更多还原