Mixture Based Outlier Filtration

Abstract

Success/failure of adaptive control algorithms – especially those designed using the Linear Quadratic Gaussian criterion – depends on the quality of the process data used for model identification. One of the most harmful types of process data corruptions are outliers, i.e. ‘wrong data’ lying far away from the range of real data. The presence of outliers in the data negatively affects an estimation of the dynamics of the system. This effect is magnified when the outliers are grouped into blocks. In this paper, we propose an algorithm for outlier detection and removal. It is based on modelling the corrupted data by a two-component probabilistic mixture. The first component of the mixture models uncorrupted process data, while the second models outliers. When the outlier component is detected to be active, a prediction from the uncorrupted data component is computed and used as a reconstruction of the observed data. The resulting reconstruction filter is compared to standard methods on simulated and real data. The filter exhibits excellent properties, especially in the case of blocks of outliers.