Key variable identification using discriminant analysis

Fault identification aims to identify key variables most relevant to diagnose a specific fault. A new fault identification approach based on the partial F -value with the cumulative percent variation (CPV) is proposed. Although the partial F -value provides the better way to interpret the single discriminant function than the fault direction and the standardized fault direction, it still suffers from the irrelevant information and low computation efficiency. To improve its identification performance and reduce the computational complexity, the CPV based on each variable’s maximum variation is proposed to determine candidate variables. These candidate variables are sufficient to express all change information of the abnormal behavior. Applying the proposed method to the Tennessee Eastman process (TEP), the results show more reliable fault identification than the fault direction, the standardized fault direction, and more efficient computation than the partial F -values.