Interval valued PCA-based approach for fault detection in complex systems

Abstract

The aim of this study is to emphasis on the detection of process sensor faults based on Principal Component Analysis (PCA). In real life case, the uncertainties of the sensor data are influencing the system and causing some difficulties in the control decision making, which in turn evokes and increases the number of false alarms and imprecise decisions. In its standard form, PCA makes no distinction between data points and the associated measurement errors which vary depending on experimental conditions. As a result, a contemporary way of representing the influence of these uncertainties on sensors has been used, namely, a representation of data in the form of interval-valued. Process modeling has been performed based on PCA for interval-valued data, where four of the most known methods have been tested. To limit the rate of false alarms, a threshold, with a certain confidence level, has been developed for both of the Hotelling’s T2, Q-statistics, and new statistics to detect the process’s faults. To confirm the ability of the proposed approach, synthetic data has been implemented, simulated, and tested on the proposed sensor fault detection. Finally, cement rotary kiln data have been tested to validate the proposed approach in reducing false alarms and missed detection rates.