A Neural Network Based Outlier Identification and Removal Scheme

Identifying and removing the outliers is important in order to make the data more trustworthy and improve the reliability of fault detection, since outliers in the measured data can cause false alarms. An online outlier identification and removal (OIR) scheme, suitable for nonlinear dynamic systems, is proposed in this paper. A neural network (NN) is utilized to estimate the actual outlier-free system states using only the measured system states which involve outliers. Outlier identification is performed online by finding the difference between measured and estimated states and comparing it with its median and standard deviation over a dynamic time window. Furthermore, the neural network weight update law is designed such that the detected outliers will not affect the state estimation. The proposed OIR scheme is then combined with fault diagnosis scheme as a preprocessing unit, in order to improve fault detection performance. A separate model-based fault detection observer is designed which uses the estimated outlier-free states to perform fault diagnosis. Finally a simple linear system is used to verify the scheme in simulations followed by a piston pump test bed study.