Dynamic process fault diagnosis using improved Fisher discriminant analysis - An approach towards IoT

The traditional fault detection methods have certain detection delay for dynamic processes with strong nonlinearity. In order to increase fault detection rate and decrease the fault detection delay, this paper proposed a new fault isolation and diagnosis method. The faulty and normal samples are separated using moving window Fisher discriminant analysis combining with mean and variance of projection error, then obtain the fault point position by hypothesis testing theory. Furthermore, the projection vector is revised by adding the auxiliary deviation. To identify the fault variables, relative error of variance is presented and compared with traditional complete deposition construction plots method. The simulation results of Tennessee Eastman benchmark process fault data sets show the advantages of this proposed method in fault isolation and diagnosis.