(345i) An Integrated Scheme for Oscillation Detection and Diagnosis from Industrial Data

Shu Xu¹, Willy Wojsznis²,
Mark Nixon³, Michael Baldea¹ and Thomas F. Edgar¹

(1) McKetta Department of Chemical
Engineering, The University of Texas at Austin, Austin, TX, (2) Innovation
Center, Emerson Process Management, Austin, TX, (3)Process Management, Emerson,
Austin, TX

As an important type of plant-wide disturbances, oscillations
generated in a single unit can propagate to several units in the plant and can negatively
affect the overall control performance of the process. Thus, it is necessary to
detect and diagnose such oscillations in three steps: (1) isolating relevant
process variables containing such oscillations; (2) diagnosing the root cause;
(3) finding the occurring time.  For step (1), the spectra envelope method
(Jiang et al.,
2007) provides an intuitive way to visualize the dominant frequencies in the
multivariate data set and a fast way to select corresponding variables
containing such frequencies so that the users no long need to perform frequency
analysis on individual variables.  For step (2), the transfer entropy
defined in Equation (1) (Schreiber, 2000) measures the
information transfer from x to  by
evaluating the reduction of uncertainty while assuming predictability(Ping et al.,
2013), and it outperforms other causality analysis methods such as the Granger's
causality (Yuan & Qin,
2014) when the process cannot be approximated by a linear model. For step
(3), the wavelet power spectrum demonstrated by Figure 1provides an intuitive
way to find the time information of the frequency change corresponding to
oscillation occurring.  In this paper, an integrated scheme is proposed,
which consists of above methods corresponding to each step: a spectral envelope
method used for identifying variables having common oscillations, a transfer
entropy method used for root cause diagnosis, and a wavelet power spectrum used
for finding the oscillation occurring time based on the root cause variable. Industrial
case studies are presented to demonstrate the proposed scheme.

where
p represents the complete or conditional probability density function, ,  ,  is the
sampling period, and h is the prediction horizon.

Figure 1 Wavelet power spectrum demonstration (Aguiar-Conraria & Soares, 2011)

(a), where   (b) wavelet power spectrum of  (c) Global wavelet power spectrum°ªaverage
wavelet power for each frequency (d) Fourier power spectral density Reference

Aguiar-Conraria, L., & Soares, M. J. (2011). The Continuous Wavelet
Transform: A Primer. Braga, Portugal: Economics Department, University of
Minho.

Jiang, H., Shoukat Choudhury, M. A. A., & Shah, S. L.
(2007). Detection and diagnosis of plant-wide oscillations from industrial data
using the spectral envelope method. Journal of Process Control, 17(2),
143-155.

Ping, D., Fan, Y., Tongwen, C., & Shah, S. L. (2013).
Direct Causality Detection via the Transfer Entropy Approach. Control
Systems Technology, IEEE Transactions on, 21(6), 2052-2066.

Schreiber, T. (2000). Measuring Information Transfer. Physical
Review Letters, 85(2), 461-464.

Yuan, T., & Qin, S. J. (2014). Root cause diagnosis of
plant-wide oscillations using Granger causality. Journal of Process Control,
24(2), 450-459.