A New DROS-Extreme Learning Machine With Differential Vector-KPCA Approach for Real-Time Fault Recognition of Nonlinear Processes

作者：Xu, Yuan; Ye, Liang-Liang; Zhu, Qun-Xiong^*

来源：Journal of Dynamic Systems Measurement and Control-Transactions of the ASME, 2015, 137(5): 051011.

DOI：10.1115/1.4028716

• dynamic recurrent OS-ELM
• differential vector-KPCA
• fault recognition

In this paper, a new dynamic recurrent online sequential-extreme learning machine (DROS-ELM) OS-ELM with differential vector-kernel based principal component analysis (DV-KPCA) fault recognition approach is proposed to reconstruct the process feature and detect the process faults for real-time nonlinear system. Toward this end, the differential vector plus KPCA is first proposed to reduce the dimension of process data and enlarge the feature difference. In DV-KPCA, the differential vector is the difference between the input sample and the common sample, which is obtained from the historical data and represents the common invariant properties of the class. The optimal feature vectors of input sample and the common sample are obtained by KPCA procedure for the difference vectors. Through the differential operation between the input vectors and the common vectors, the reconstructed feature is derived by calculating the two-norm distance for the result of differential operation. The reconstructed features are then utilized to detect the process faults that may occur. In order to enhance the accuracy of fault recognition, a new DROS-ELM is developed by adding a self-feedback unit from the output of hidden layer to the input of hidden layer to record the sequential information. In the DROS-ELM, the output weight of feedback layer is updated dynamically by the change rate of output of the hidden layer. The DV-KPCA for feature reconstruction is exemplified using UCI handwriting (UCI handwriting recognition data: Database, using "Pen-Based Recognition of Handwritten Digits" produced in the Department of Computer Engineering Bogazici University, Istanbul 80815, Turkey, 1998), which the classification accuracy is obviously enhanced. Meanwhile, the DROS-ELM for process prediction is tested by the sunspot data from 1700 to 1987, which also shows better prediction accuracy than common methods. Finally, the new joint DROS-ELM with DV-KPCA method is exemplified in the complicated Tennessee Eastman (TE) benchmark process to illustrate the efficiencies. The results show that the DROS-ELM with DV-KPCA shows superiority not only in detection sensitivity and stability but also in timely fault recognition.

• 出版日期2015-5
• 单位北京化工大学