Fault isolation of non-Gaussian processes based on reconstruction

Abstract

Classical fault detection method has been successfully applied to practical industrial processes. However, fault isolation is still a difficult issue yet to be solved. It is obvious because the existing fault isolation methods have not fully used the embedded information of the concerned faults, where accuracy and reliability of fault isolation cannot be guaranteed. For this purpose this paper presents a novel data driven fault isolation approach for non-Gaussian processes. The proposed method firstly utilizes an offline learning mechanism to obtain information on normal operating conditions together with the learning on the variance and covariance of the faults. And then it is followed by the retrieving of the fault relevant directions and the construction of detection statistics together with the relevant confidence limit. As such, the fault identification model based on reconstruction for fault and residual subspace is obtained which represents healthy operation status of the process. In this way an effective data driven fault isolation algorithm is established. In addition, the proposed method has been applied to the fault isolation for an electro-fused magnesia furnace (EFMF) and desired results have been obtained.