Sensor Fault Diagnosis Method Based On Data Analysis

Research work on sensors fault detection and diagnosis method has been studied and deployed. All methods were test in simulated CSTR model. The detection and diagnosis results have been analyzed. Completed work is summarized as following:The paper gives a integrated research based on PCA from fault detection, fault diagnosis, reconstruction fault to a new fault detection method based on KPCA. It illuminated the shortcoming of contribution plot and afforded a new way of Analogical Degree. Contrasting to contribution plot, the new way can differentiate the fault sensor on line by computing cosine value of angle between real time measure data and history fault knowledge. It also proved that the KPCA is more predominant than PCA in extracting the nonlinear feature of fault signals.ICA was demonstrated to separate independent components and extract feature, in order to solve the non-Gaussian characteristic problem of data. The fault detection was implemented by establishing three statistical parameters I 2, I e2and SPE . The confidence bounds were determined by kernel density estimation. The fault detection was implemented by contribution plot. Considering complex nonlinear relations between data, the KICA was put forward for fault detection. The simulated result proved that the KICA is superior to ICA in fault detection.Fisher discriminant analysis was applied in sensors fault diagnosis. Information between fault classes was considered which shows that FDA is more proficient than PCA for diagnosis faults.