Machine learning technique for data-driven fault detection of nonlinear processes

This paper proposes a new machine learning method for fault detection using a reduced kernel partial least squares (RKPLS), in static and online forms, for handling nonlinear dynamic systems. The choice of the fault detection method has a vital role to improve efficiency and safety as well as production. The kernel partial least squares is a nonlinear extension of partial least squares. The present method has been mostly used as a monitoring method for nonlinear processes. Thus, the standard method cannot perform properly and quickly when the training data set is large. The main contributions of the suggested approach are: the approximation of the components retained by the standard method and the reduction in the computation time as well as the false alarm rate. Using the reduced principal, the online suggested method is presented for fault detection of nonlinear dynamic processes. The online reduced method is developed to monitor the dynamic process online and update the reduced reference model. For this reason, the moving window RKPLS is proposed. The general principle is to check if the new useful observation satisfies, in the feature space, the condition of independencies between variables. Thereafter, the relevance of the suggested methods is used to monitor the chemical stirred tank reactor benchmark process, the air quality and the tennessee eastman process. The simulation results of the suggested methods are compared to the standard one.