CIESC Journal ›› 2019, Vol. 70 ›› Issue (12): 4770-4776.DOI: 10.11949/0438-1157.20191350

• Process system engineering • Previous Articles     Next Articles

Features extracted from auto-encoder based echo state network and its applications to process modeling

Bao ZHU1,2(),Junfei QIAO1()   

  1. 1. Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
    2. POWERCHINA Resources Ltd. , Beijing 100048, China
  • Received:2019-11-05 Revised:2019-11-15 Online:2019-12-05 Published:2019-12-05
  • Contact: Junfei QIAO

基于自编码神经网络特征提取的回声状态网络研究及过程建模应用

朱宝1,2(),乔俊飞1()   

  1. 1. 北京工业大学信息学部,北京 100124
    2. 中国电建集团海外投资有限公司,北京 100048
  • 通讯作者: 乔俊飞
  • 作者简介:朱宝(1987—),男,博士,zhubao@powerchina.cn
  • 基金资助:
    国家自然科学基金重大项目(61890930)

Abstract:

In the process of complex chemical modeling, the traditional static neural network modeling can not meet certain accuracy because of the time sequence, high nonlinearity and high dimension of process data. To solve this problem, a feature extracted from auto-encoder based echo state network (FEAE-ESN) is proposed in this paper. In the traditional echo state network (ESN) method, the number of nodes in the reserve pool of ESN is large, and then the dimension of the reserve pool output is very high. Therefore, to solve this problem, the auto-encoder is used to extract features from the reserve pool output of well-trained ESN. Through the feature extraction of auto-encoder, on one hand, the dimension of the output of the reserve pool can be effectively reduced, thereby reducing the complexity of the data; on the other hand, the collinearity of the outputs of the original reserve pool can be removed through extracting features, which can further improve the calculation performance of generalized inverse. Ultimately, the modeling accuracy of ESN is improved. The proposed FEAE-ESN is applied to modeling the Tennessee-Eastman process. The simulation results verify the effectiveness of the proposed method.

Key words: auto-encoder, echo state network, feature extraction, soft sensor, process modeling

摘要:

在复杂化工建模过程中,由于过程数据的时序性、高非线性以及高维数的特点,导致传统的静态神经网络建模无法满足一定的精度。为了解决该问题,提出一种基于自编码神经网络特征提取的回声状态网络模型(features extracted from auto-encoder based echo state network,FEAE-ESN)。传统回声状态网络(echo state network, ESN)方法中,储备池的节点数目很多,输出的维数很高,数据间存在共线性。为解决上述问题,待回声状态网络训练好之后,使用自编码神经网络对其储备池输出进行特征提取。通过自编码网络特征提取,一方面可以有效地降低储备池输出的维数,从而降低数据的复杂度;另一方面提取的特征去除了原有储备池输出的共线性,能够进一步提高广义逆的计算性能;最终提高回声状态网络的建模精度。所提方法FEAE-ESN用于田纳西-伊斯曼复杂过程建模,仿真结果验证了所提方法的有效性。

关键词: 自编码神经网络, 回声状态网络, 特征提取, 软测量, 过程建模

CLC Number: