【作者】 刘颖；

【导师】 宋执环；

【作者基本信息】 浙江大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 计数数据是一种重要数据类型,广泛存在于众多学科与工程领域。计数数据的观测值以非负整数的形式表示,属离散变量。其非负整数和离散特征,决定了有必要建立离散计数数据回归模型,以探究影响事件发生率的关键因素和实时预报事件的发生次数。工业过程正朝着大规模、复杂化的方向发展,所产生的测量数据呈指数级增长,以过程变量预测质量变量的软测量技术应运而生。目前计数数据建模在工业领域的应用尚不成熟,将传统的软测量方法运用到工业计数数据建模时仍具有一些局限;同时工业过程数据往往具有高维、非线性、非高斯等特性,传统的计数数据建模与分析方法面临着信息挖掘不充分的缺陷。所以建立工业计数数据模型有较多待解决的问题。本文结合计数数据特性,借助深度学习具有强大的特征提取能力,基于自编码器对工业计数数据进行质量预测建模研究,主要研究内容如下:(1)针对常规计数数据模型特征提取能力弱的问题,提出了基于深度特征提取器的泊松回归模型。将计数数据泊松回归引入深度堆叠降噪自编码器,有效提取能反映原始输入数据的深度特征,提高了模型的特征挖掘能力,提升了计数数据回归模型的预测效果,实现更好的工业质量预测。(2)针对常规自编码器不能提取质量变量相关特征的问题,提出了基于监督泊松自编码器的计数数据回归模型。在预训练的解码阶段引入质量变量来指导特征提取,通过泊松网络层将质量变量计数数据集成到深度堆叠编码器结构中,使得模型学习到的特征表示与计数数据类型的质量变量高度相关,提升了特征提取效率,并且提升了计数质量数据的预测效果。(3)针对常规自编码器特征利用不充分的问题,在监督泊松自编码器的基础上提出了基于注意力堆叠监督泊松自编码器的计数数据回归模型。通过引入注意力机制来衡量各个隐层特征对于计数数据输出值的贡献,并用泊松网络层来集成隐层特征,运用蕴含在不同隐层特征中的信息共同预测计数数据质量变量,建立了输入到输出的多条连接通道,提升了隐层特征的有效利用率,且改善了模型的预测性能。更多还原

【Abstract】 Counting data is an important data type,which widely exists in many disciplines and engineering fields.Observed values of count data are expressed in the form of nonnegative integers,which are discrete variables.Because of its characteristics of nonnegative and discrete,it is necessary to establish a regression model of discrete count data to explore the key factors affecting the incidence of events and predict the number of forecast events in real time.Industrial processes are developing in the direction of large-scale and complexity,and the measurement data generated is increasing exponentially.The soft sensor technology which predicts quality variables by process variables is widely used.At present,the application of counting data modeling in the industrial field is still immature.Traditional soft sensor methods have certain limitations when applied to industrial counting data modeling.At the same time,the industrial process data often has the characteristics of high dimension,nonlinear and non-Gaussian,and traditional counting data modeling and analysis methods are faced with the defect of insufficient information mining.Therefore,there are many problems to be solved in the establishment of industrial counting data model.In this paper,combined with the characteristics of count data,with the help of deep learning with powerful feature extraction capabilities,the quality prediction modeling research of industrial count data based on autoencoders is carried out.The main research contents are as follows:(1)Aiming at the problem of weak feature extraction capabilities of conventional counting data model,a Poisson regression model based on deep feature extractor is proposed.By introducing a deep stacking noise reduction autoencoder,the information contained in the input data is fully mined,and the depth features that can reflect the original input are effectively extracted,which improves the feature mining ability of the model.Combined with the Poisson regression method for counting data,the prediction effect of the model is improved,and better industrial quality prediction is achieved.(2)Aiming at the problem that conventional autoencoders cannot extract features related to quality variables,a count data regression model based on supervised Poisson autoencoders is proposed.By introducing quality variables in the pre-training decoding stage to guide feature extraction,and integrating them into the deep stacked encoder structure through the Poisson regression network layer for count data,the feature representations learned by the model are highly correlated with the count data quality variables.The efficiency of feature extraction is improved,and the effect of counting data prediction is improved.(3)Aiming at the problem of insufficient utilization of the features of conventional autoencoders,a count data regression model based on the attention stacked supervised Poisson autoencoder is proposed on the basis of the supervised Poisson autoencoder.The attention mechanism is used to measure the contribution of each hidden layer feature to output value of the count data,and the Poisson network layer is used to integrate the potential features.The information contained in the different hidden layer features is used to jointly predict the count data.By establishing multiple connection channels from input to output,the effective utilization of hidden layer features is improved,and the performance of the model is improved.更多还原