基于双阈值AdaBoost算法的4-CBA含量软测量建模

发布时间：2018-05-16 04:28

  本文选题：AdaBoost算法 + 软测量　； 参考：《化工学报》2017年05期

【摘要】：针对PX氧化过程中4-CBA含量无法在线测量的问题,提出了一种基于双阈值更新样本权重的AdaBoost算法,该算法以BP神经网络作为弱学习器,采用轮盘赌方法根据样本权重在训练样本集中选择部分样本训练弱学习器,采用上一轮弱学习器的训练相对误差绝对值来更新所有训练样本的权重,在此基础上,用双阈值对样本误差范围进行划分,然后用不同的权重因子与原来的样本权值相乘实现样本权值的二次更新。该过程降低了含有大误差的样本的权值,增加了较大误差的样本的权值,从而减小了在下一轮训练过程中选到异常样本的概率。分别采用5种不同的方法并用实测的工业数据建立了4-CBA含量软测量模型,仿真结果表明用提出的改进AdaBoost算法建立的4-CBA含量软测量模型,其预测误差小于其他方法建立的模型误差。
[Abstract]:In order to solve the problem that 4-CBA content can not be measured on line during PX oxidation, a new AdaBoost algorithm based on double threshold to update the weight of samples is proposed. The BP neural network is used as a weak learner in this algorithm. The roulette method is used to select part of the training samples to train the weak learner according to the weight of the samples, and the absolute value of the relative error of the last round of the weak learner is used to update the weight of all the training samples. The range of sample error is divided by double threshold, and then the quadratic updating of sample weight is realized by multiplying different weight factors with the original sample weight. This process reduces the weight of samples with large errors and increases the weights of samples with large errors, thus reducing the probability of selecting abnormal samples in the next round of training. The soft sensing model of 4-CBA content is established by using five different methods and the measured industrial data. The simulation results show that the proposed improved AdaBoost algorithm is used to establish the soft sensor model of 4-CBA content. The prediction error is smaller than the model error established by other methods.
【作者单位】： 南京邮电大学自动化学院;
【基金】：国家自然科学基金项目(61203213)~~
【分类号】：TP183;TQ245.12
，

本文编号：1895463