Online Fault Detection Based on OCKELM and Sparse Dictionary

In order to improve the accuracy and to curb the matrix expansion of the kernel learning methods in the process of online fault detection, a new method based on one-class kernel extreme learning machine (OCKELM) and incremental learning is proposed. Firstly, the OCKELM fault detection target function is established only depended on the target samples. And the expressions of kernel function and kernel weight vector are derived. Sparse dictionary is introduced to prevent the expansion of the dictionary scale. During the process of forward and backward sparse, with the strategy of construction and pruning, a set of key nodes with a predetermined scale is selected based on minimizing the cumulative coherence of the online dictionary. Besides, according to the node selection results, the model parameters are updated online recursively by using the elementary transformation of the matrix and the block matrix inverse formula. The proposed method is applied to the artificial data set and UCI data set. The experimental results show that the testing time consumption of the proposed method is at the level of 10-4s and the online detection can be realized. Moreover, compared with the existing mature offline fault detection methods, this method performs better in F1, AUC, G-mean and fault detection rate.