Dynamic sensor fault detection approach using data-driven techniques

Sensor fault detection is an important phase for process surveillance. Indeed, successful execution of process tasks depends on the state of the available data. In industrial applications, systems have an uncertain behavior, so methods based on interval principles are useful in this context, such as interval kernel principal component analysis (IKPCA) and Interval-Kernel-Partial-Least Square (IKPLS). These techniques do not always achieve efficiency since they are based on an invariant time. To overcome this restriction, we used a sliding windows principle to develop these methods in dynamic uncertain systems. In addition, a small size of the sliding window can improve the computation time and adapt to the rapidly changing process dynamics. It is for this reason that we propose a second method; the reduced rank moving window IKPCA method (MW-RRIKPCA), In order to improve the performance of the suggested methods based on sliding windows, compared to classical ones, we carried out a comparative study using Tennessee Eastman process (TEP) data and the air quality monitoring network (AIRLOR).