基于改进Bagging算法的高斯过程集成软测量建模

为提高对工况复杂的工业过程进行软测量建模的模型精度和泛化能力,提出了一种基于改进Bagging算法的高斯过程集成软测量建模方法。该算法采用高斯过程回归算法建立集成学习模型的基学习器,并在Bagging算法对训练样本重采样生成基学习器训练子集的基础上,采用基于正则化互信息的特征排序指标进行基学习器的输入特征抽取,实现有监督的特征扰动,从而改善学习器的差异度。待测样本进行软测量估计时,根据各高斯过程基学习器输出的方差自适应地选择基学习器进行集成输出。采用工业双酚A生产装置反应器的现场数据建模仿真,结果表明该方法是有效的。     

一种新颖的小样本整体趋势扩散技术

基于数据驱动的生产过程建模、优化与控制是当今学术界与企业界的研究与应用热点。大数据时代小样本问题不可忽视。针对诸如人工神经网络(ANNs)、极限学习机(ELMs)等传统建模方法在小样本条件下难以获得较高的学习精度,提出了一种新颖的多分布整体趋势扩散技术(multi-distribution mega-trend-diffusion, MD-MTD)用于提升小样本学习精度。通过整体扩散技术推估小样本属性可接受范围,在整体趋势扩散的基础上,增加了均匀分布和三角分布描述小样本数据特性,生成虚拟样本,填补小样本数据点间的信息间隔。利用标准函数产生标准样本,在正交实验和不均匀样本实验下论证了MD-MTD的合理性和有效性,用MLCC和PTA两个实际的工业数据集进一步验证了MD-MTD的实用性。实验结果表明,MD-MTD能提高小样本学习精度8%以上。     

城市固废炉排炉焚烧过程二英排放浓度数值仿真

城市固废焚烧（MSWI）过程产生的二英（DXN）是至今机理仍复杂不清的剧毒污染物，获悉DXN在炉排炉内的生成、燃烧和再生成等过程的边界条件对降低污染排放极为重要。对此，本文提出了城市固废炉排炉焚烧过程DXN排放浓度数值仿真方法。首先，依据面向DXN的典型炉排炉MSWI工艺流程，描述焚烧炉内固相燃烧、气相燃烧、高温换热和低温换热等与DXN相关反应的机理。接着，依据上述所划分区域，结合实际MSWI过程相关参数构建DXN数值仿真模型。最后，基于烟气分流分率所表征的反应物浓度和不同区域的反应温度进行单因素分析，以获取G1处DXN浓度的边界条件，并基于正交实验分析分流分率和反应温度对G1处DXN浓度的影响，进而获得最优参数组合。基于北京某MSWI电厂实际数据的数值仿真分析与验证，表明了该数值仿真模型的有效性，为后续优化控制G1处的DXN排放浓度提供了支撑。     

基于FEEMD-AE与反馈极限学习机组合模型预测研究与应用

针对现有工业过程非平稳时间序列中的特征提取及预测问题,提出了基于快速集合经验模态分解（fast ensemble empirical mode decomposition,FEEMD）、近似熵（approximate entropy,AE）和反馈极限学习机（feedback extreme learning machine,FELM）的组合模型。首先,针对复杂非平稳时序数据,采用FEEMD方法将其分解为从高频到低频的相对平稳的本征模态函数分量和余项;其次,为解决经过FEEMD分解出来的分量复杂度问题,运用近似熵（AE）计算分量复杂度并进行特征重构,以降低分量复杂性;然后,基于传统ELM结构,通过引入反馈机制,在输出层与隐含层之间增加反馈层用来记忆隐含层输出数据,并计算数据趋势变化率动态更新反馈层输出,形成反馈极限学习机（FELM）,对非线性动态系统的下一时刻输出进行预测;最后,将所提出的组合预测模型通过UCI标准数据集与精对苯二甲酸（PTA）溶剂系统进行建模仿真,仿真结果表明,提出的组合模型预测方法能够得到较高的预测精度,为实际生产操作优化提供了一定的指导。     

基于选择性集成核学习算法的固废焚烧过程二噁英排放浓度软测量

城市固废焚烧(MSWI)过程排放的二噁英(DXN)是被称为"世纪之毒"的持续性污染物。该过程的多阶段、多温度区间的物理化学特性导致DXN排放浓度的机理模型难以构建。工业实际中通常以月或季为周期耗时近1周时间在实验室以离线化验方式滞后检测。针对这些问题,提出了基于选择性集成(SEN)核学习算法的DXN排放浓度软测量方法。首先,基于先验知识给出候选核参数集和候选惩罚参数集,采用核学习算法构建基于这些超参数的候选子子模型;然后,耦合优化和加权算法对相同核参数的候选子子模型进行选择与合并,进而得到基于不同核参数的候选SEN子模型集合;最后,再次采用优化和加权算法获得结构与超参数自适应的多层SEN软测量模型。采用UCI平台水泥抗压强度和焚烧过程DXN数据验证了所提方法的有效性。     

基于选择性集成核学习算法的固废焚烧过程二噁英排放浓度软测量

城市固废焚烧(MSWI)过程排放的二噁英 (DXN)是被称为“世纪之毒”的持续性污染物。该过程的多阶段、多温度区间的物理化学特性导致DXN排放浓度的机理模型难以构建。工业实际中通常以月或季为周期耗时近1周时间在实验室以离线化验方式滞后检测。针对这些问题,提出了基于选择性集成(SEN)核学习算法的DXN排放浓度软测量方法。首先,基于先验知识给出候选核参数集和候选惩罚参数集,采用核学习算法构建基于这些超参数的候选子子模型;然后,耦合优化和加权算法对相同核参数的候选子子模型进行选择与合并,进而得到基于不同核参数的候选SEN子模型集合;最后,再次采用优化和加权算法获得结构与超参数自适应的多层SEN软测量模型。采用UCI平台水泥抗压强度和焚烧过程DXN数据验证了所提方法的有效性。     

基于FEEMD-AE与反馈极限学习机组合模型预测研究与应用

针对现有工业过程非平稳时间序列中的特征提取及预测问题,提出了基于快速集合经验模态分解(fast ensemble empirical mode decomposition,FEEMD)、近似熵(approximate entropy,AE)和反馈极限学习机(feedback extreme learning machine,FELM)的组合模型。首先,针对复杂非平稳时序数据,采用FEEMD方法将其分解为从高频到低频的相对平稳的本征模态函数分量和余项;其次,为解决经过FEEMD分解出来的分量复杂度问题,运用近似熵(AE)计算分量复杂度并进行特征重构,以降低分量复杂性;然后,基于传统ELM结构,通过引入反馈机制,在输出层与隐含层之间增加反馈层用来记忆隐含层输出数据,并计算数据趋势变化率动态更新反馈层输出,形成反馈极限学习机(FELM),对非线性动态系统的下一时刻输出进行预测;最后,将所提出的组合预测模型通过UCI标准数据集与精对苯二甲酸(PTA)溶剂系统进行建模仿真,仿真结果表明,提出的组合模型预测方法能够得到较高的预测精度,为实际生产操作优化提供了一定的指导。     

基于流形正则化半监督学习的污水处理操作工况识别方法

污水处理过程容易受外界冲激扰动影响,引发污泥上浮、老化、中毒、膨胀等故障工况,导致出水水质质量差,能源消耗高等问题,如何快速准确识别污水操作工况故障至关重要。针对污水工况识别过程中现有监督学习方法未利用大量未标记数据蕴含的丰富操作工况信息,采用基于流形正则化极限学习机的半监督学习方法,监视生化污水处理过程操作运行工况。该方法在学习过程中,在标记和未标记数据输入空间构建图拉普拉斯算子,通过随机特征映射建立隐含层,在流形正则化框架下,求解隐含层和输出层之间的权重,保留随机神经网络的计算效率和泛化性能。仿真实验结果表明,基于半监督极限学习机的污水处理工况识别在准确率与可靠性方面相对优于基本极限学习机方法。     

基于流形正则化半监督学习的污水处理操作工况识别方法

污水处理过程容易受外界冲激扰动影响,引发污泥上浮、老化、中毒、膨胀等故障工况,导致出水水质质量差,能源消耗高等问题,如何快速准确识别污水操作工况故障至关重要。针对污水工况识别过程中现有监督学习方法未利用大量未标记数据蕴含的丰富操作工况信息,采用基于流形正则化极限学习机的半监督学习方法,监视生化污水处理过程操作运行工况。该方法在学习过程中,在标记和未标记数据输入空间构建图拉普拉斯算子,通过随机特征映射建立隐含层,在流形正则化框架下,求解隐含层和输出层之间的权重,保留随机神经网络的计算效率和泛化性能。仿真实验结果表明,基于半监督极限学习机的污水处理工况识别在准确率与可靠性方面相对优于基本极限学习机方法。     

基于EGK'M-RBF网络的顺丁橡胶门尼黏度预测

提出一种基于增强的全局K'-means算法(EGK'M)-RBF网络的建模方法,该方法采用作者提出的EGK'M来确定RBF网络隐含层的结构,包括隐含层中心个数、中心位置以及隐含层扩展常数,采用KPCA提取非线性特征信息,实现辅助变量的二次选择.并与基于PCA和EGK'M-RBF网络模型、基于KPCA和K-means算法...     

Virtual sample generation method based on improved megatrend diffusion and hidden layer interpolation with its application

The time and economic cost of obtaining difficult-to-measure quality or environmental pollution indices data for complex industrial processes are very high, which leads to the scarcity of labeled modeling samples. Aimed at this problem, a new virtual sample generation method based on improved megatrend diffusion and hidden layer interpolation is proposed. It is applied to the dioxin (DXN) emissions prediction of municipal solid waste cineration process. First, the true sample input/output sample space is expanded by using improved megatrend diffusion technology (MTD) based on the sub-regional Euclidean distance. Next, the virtual sample input is generated using an equal interval interpolation method, and the virtual sample output is obtained by combining the mapping model and the pruning mechanism. Then, the hidden layer interpolation method based on the improved random weight neural network with regularization mechanism is used to obtain the virtual sample output and input, and the virtual sample is deleted by combining with the expansion space. Finally, the above-mentioned complementary input/output virtual samples are mixed with the original true samples to realize the expansion of the modeling data capacity. The validity and rationality of the proposed method are verified by benchmark data set and industrial process DXN data.     

Unification of neural and statistical modeling methods that combine inputs by linear projection

Empirical modeling methods that combine inputs by linear projection include linear methods such as, ordinary least-squares regression, partial least-squares regression, principal components regression, and nonlinear methods such as, backpropagation networks with a single hidden layer, projection pursuit regression, nonlinear partial least-squares regression, and nonlinear principal components regression. In this paper, these popular modeling techniques are unified to yield a single method called nonlinear continuum regression (NLCR). This unification is based on the insight provided by a common framework for empirical modeling methods, and is achieved by using activation functions that adapt to the measured data, a common optimization criterion for finding the projection directions, and a hierarchical training methodology that allows efficient modeling. The adaptive-shape activation functions are determined by univariate smoothing in the space of the projected input versus output. The NLCR optimization criterion contains an adjustable parameter that controls the degree of overfitting or bias of the model, and spans the continuum of methods from projection pursuit regression or backpropagation networks to nonlinear principal components regression. Consequently, NLCR results in models that are usually more general and compact than those obtained by existing methods based on linear projection, while eliminating the need for arbitrary selection of an empirical modeling method based on linear projection for a given task. The improved modeling ability of NLCR and its performance on different types of training data are illustrated by examples based on simulated and industrial data.     

利用人工神经网络预测复相陶瓷材料组分含量的研究

根据人工神经网络（ANN）的BP(back propagation）算法,预测复相陶瓷各组分的体积分数的神经网络模型,模型由三层神经元组成,分别为输入层、隐含层和输出层用以模拟人脑的结构,输入层参数由两部分组成,一部分为抗弯强度、硬度和断裂韧性等力学性能,另一部分包括相应各组分的弹性模量和热膨胀系数,以用来辨识不同的材料系统,输出层参数是复相陶瓷中各组分的体积分数,只要训练样本值足够精确,预测模型就能够预测已有的陶瓷系统的组分含量,同时,模型能够预估未知材料系统的组分含量。计算证明,模型的容错性较好,因此对开发新型复相陶瓷非常有益。     

Input variable scaling for statistical modeling

Input variable scaling is one of the most important steps in statistical modeling. However, it has not been actively investigated, and autoscaling is mostly used. This paper proposes two input variable scaling methods for improving the accuracy of soft sensors. One method statistically derives the input variable scaling factors; the other one uses spectroscopic data of a material whose content is estimated by the soft sensor. The proposed methods can determine the scales of the input variables based on their importance in output estimation. Thus, it can reduce the negative effects of input variables which are not related to an output variable. The effectiveness of the proposed methods was confirmed through a numerical example and industrial applications to a pharmaceutical and a distillation processes. In the industrial applications, the proposed methods improved the estimation accuracy by up to 63% compared to conventional methods such as autoscaling with input variable selection.     

基于循环自动编码器的间歇过程故障监测

针对间歇过程数据非线性、动态性特征,提出一种基于循环自动编码器(recurrent autoencoder,RAE)的过程故障监测方法。采用长短时记忆(long short-term memory,LSTM)循环神经网络构建自动编码器建立监控模型,相比传统自动编码器,其能有效挖掘时序样本间的动态关联信息。该方法首先利用批次展开与变量展开相结合的三步展开方法将间歇过程数据展开成二维,并通过滑动窗采样得到模型输入序列;然后使用LSTM构建自动编码器,重构输入序列。进一步,利用重构误差构造平方预测误差(squared prediction error, SPE)统计量实现在线监测。最后将所提方法应用于青霉素发酵仿真和重组大肠杆菌发酵过程监测,结果表明,该方法能及时监测到故障,具有较好的监测性能。     

Mixture semisupervised principal component regression model and soft sensor application

Traditionally, data‐based soft sensors are constructed upon the labeled historical dataset which contains equal numbers of input and output data samples. While it is easy to obtain input variables such as temperature, pressure, and flow rate in the chemical process, the output variables, which correspond to quality/key property variables, are much more difficult to obtain. Therefore, we may only have a small number of output data samples, and have much more input data samples. In this article, a mixture form of the semisupervised probabilistic principal component regression model is proposed for soft sensor application, which can efficiently incorporate the unlabeled data information from different operation modes. Compared to the total supervised method, both modeling efficiency and soft sensing performance are improved with the inclusion of additional unlabeled data samples. Two case studies are provided to evaluate the feasibility and efficiency of the new method. © 2013 American Institute of Chemical Engineers AIChE J 60: 533–545, 2014     

Semisupervised Bayesian method for soft sensor modeling with unlabeled data samples

Most traditional soft sensors are built upon the labeled dataset that contains equal numbers of input and output data samples. However, the output variables that correspond to quality variables and other important controlled variables are always difficult to obtain in chemical processes. Therefore, we may only obtain the output data for a small portion of the whole dataset and have much more input data samples. In this article, a semisupervised method is proposed for soft sensor modeling, which can successfully incorporate the unlabeled data information. To determine the effective dimensionality of the latent space, the Bayesian regularization method is introduced into the semisupervised model structure. Two industrial application case studies are provided to evaluate the feasibility and efficiency of the newly developed probabilistic soft sensor. © 2010 American Institute of Chemical Engineers AIChE J, 2011     

A novel imbalanced fault diagnosis method based on area identification conditional generative adversarial networks

As a vital technology for ensuring the stable operation of industrial equipment, fault diagnosis has received a lot of research in recent years. Most complex industrial processes are in normal working conditions during operation, so the amount of data collected under normal working conditions is much larger than that under fault working conditions. The uneven number of samples will lead to the imbalance of datasets and make it a challenging task to assure the overall accuracy. To address the issue, an innovative imbalanced fault diagnostic approach based on area identification conditional generative adversarial networks (AICGAN) is proposed. First, considering the imbalance between normal data (majority data) and fault data (minority data), a hybrid data generation method combining over-sampling and AICGAN generator is proposed, which effectively extends the limited minority data and overcomes the inclination to majority data to some extent. On one hand, the over-sampling algorithm reduces the impact of dataset imbalance on the AICGAN training process by linear interpolation. On the other hand, the trainable generator can create samples similar to real samples by learning the generation principle so as to enrich the minority data information and reduce the sample stacking caused by linear synthesis. The two sample production methods complement each other. Combining the raw samples, over-sampled samples, and samples generated by generator, a new dataset is constructed. Second, the new dataset is used to train the AICGAN discriminator. In addition, in order to generate samples with higher value, an auxiliary discrimination layer is added to the discriminator to control the pattern of generated samples. Third, the balanced dataset containing the linear synthesis samples and the samples generated by the trained generator are put into the classifier to obtain the fault diagnosis. The effectiveness of the proposed approach for fault diagnosis based on AICGAN is verified using the three-phase flow facility (TFF) dataset and the Tennessee Eastman (TE) dataset. The experimental results demonstrate that the AICGAN-based fault diagnosis method achieves high F1 scores on the imbalanced dataset.