Novelty Detection for Multi-Mode Process Monitoring based on Subspace Selection

Monitoring a complex system and identifying its operating modes and potential faults has been an active research subject over the last decades. A special case arises in the monitoring of multi-mode systems, where data gathered from multiple distributed sensors do not represent unequivocally the mode the system is operating in. In such scenarios, the sensors data can represent high-dimensional distribution of severe overlapping clusters. We propose a Statistical Process Control (SPC) framework that aims at dealing with the above-mentioned challenges. The proposed schema is based on randomly selected subsets of sensors combined with Bayesian decision theory. As a special use-case of multi-mode systems, we apply our framework to data gathered from Metrology devices in the semiconductor industry. The outcome of the monitoring scheme is the identification of a new fault as a new operation mode of the system. We show that the use of combined subsets of sensors along with probabilistic modeling has good potential for the monitoring of such multi-mode systems.