Building Multivariate Models from Compressed Data

The effect of data compression on multivariate analysis, especially on principal components analysis (PCA) based modeling, is evaluated in this paper. A comparative study between the “swinging door” and “wavelet compression” algorithms is performed in the context of multivariate data analysis. It is demonstrated that wavelet compression preserves the correlation between different variables better than swinging door compression. It is also demonstrated that the impact of compression increases as the dynamics of the processes become faster and more stochastic in nature. Instead of interpolation based reconstruction of swinging-door-compressed data and subsequent modeling, an iterative missing data technique is suggested for building a PCA model from swinging-door-compressed data. The performance of the proposed methodology is demonstrated using a simulated flow-network system and an industrial data set.