On-Line Multiscale Filtering of Random and Gross Errors without Process Models

Mohamed N. Nounou and Bhavik R. Bakshi

Department of Chemical Engineering
The Ohio State University, Columbus, OH 43210, USA


Abstract

Data Rectification by univariate filtering is popular for processes that lack an accurate model.  Linear filters are most popular for on-line filtering. Unfortunately, these methods are single  scale in nature and are best for rectifying data containing features and noise that are at the same  resolution in time and frequency. Consequently, for multiscale data, linear filters are forced to trade-off the extent of noise removal with the accuracy of the features retained. In contrast,
nonlinear filtering methods such as, finite impulse response median hybrid filters, and wavelet thresholding are multiscale in nature, but they can not be used for on-line rectification. This paper presents a technique for on-line nonlinear filtering based on wavelet thresholding. On-line multiscale (OLMS) rectification applies wavelet thresholding to data in a moving window of dyadic length to remove random errors. Gross errors are removed by combining wavelet
thresholding with multiscale median filtering. Theoretical analysis shows that OLMS rectification using Haar wavelets subsumes mean filters of dyadic length, while rectification with smoother boundary corrected wavelets is analogous to adaptive exponential smoothing. If the rectified measurements are not needed on-line, the quality of rectification can be further improved by averaging the rectified signals in each window. The resulting approach overcomes the boundary effects encountered in translation invariant (TI) rectification of Coifman and Donoho (1995), and is called boundary corrected translation invariant (BCTI) rectification. Examples based on synthetic and industrial data demonstrate the benefits of the on-line multiscale and boundary corrected translation invariant rectification methods.