Detection And Diagnosis Of Oscillations In Process Industry Using Time-frequency Methods

The increased scale and integration of modern industrial processes have brought considerable challenges to automatic control systems.With the implementation of the "Made in China 2025"plan,many companies have improved their standards on productivity,profitability,energy con-sumption,quality and safety.In the meantime,researchers have recognized that the low control performance of process industries is a widespread and increasingly serious problem,where oscil-lation is one of the most common abnormal phenomena.In a large industrial plant,the number of control loops can be in range of thousands and this level of automation necessitates a reliable mon-itoring mechanism to ensure the safe and efficient operation of each production unit.However,recent developments on oscillation-related techniques mainly require the restrictive assumptions that the oscillation should be stationary or linear,and its magnitude/frequency does not change too frequently.Since 1998,the empirical mode decomposition(EMD)related time-frequency tools have been extensively studied and widely applied in numerous engineering applications.Com-pared with Fourier-base transform,EMD is a major breakthrough over the linear and stationary spectral analysis.Based on the above statements,this thesis has provided following contributions in cooperation with the modern time-frequency tools to address part of the difficulties encountered in oscillation detection and diagnosis.· A local mean decomposition(LMD)based method for detecting the multiple and time-varying oscillations from nonstationary processes is proposed.The original LMD is first improved in terms of reduced end effect,adaptive window size and optimized stoppage criterion,then a robust Lempel-Ziv complexity based statistic is developed to monitor the decomposed product functions.Secondly,motivated by the fact that it is still an open prob-lem to design a real-time monitor which is capable of detecting multiple oscillations with signal intermittency and nonlinear/nonstationary properties,this work proposes the frame-work of intrinsic time-scale decomposition(ITD)with robust zero-crossing intervals(ZCIs)clustering.An effective online updating statistic is also presented for the purpose of real-time oscillation detection· Aiming at addressing the high computational load,high sampling rate and over-decomposition problem of multivariate empirical mode decomposition(MEMD),this thesis presents the fast MEMD(FMEMD)and the noise-assisted FMEMD(NA-FMEMD)algorithms.FMEMD is a computationally less-expensive alternative to MEMD that operates by applying univari-ate empirical mode decomposition(EMD)on projected signals to obtain a set of intrinsic mode functions(IMFs),which are combined with their corresponding direction vectors and then solved by least square algorithm to yield multivariate IMFs.Additionally,the noise-assisted version(NA-FMEMD)has enhanced the performance of FMEMD by imposing the inherent quasi-filter bank structure of FMEMD on the input data.Such novel method is applicable for both single-loop and plant-wide oscillations monitoring.· Inspired from the fact that only few time-frequency tools are available to process the multi-channel data,two multivariate extensions of the standard ITD algorithm are proposed to cater for a more general extract and cluster of common oscillations from multiple control loops.Similar to the FMEMD algorithm,the indirect multivariate ITD(IMITD)is general-ized in a subtle way,by making full use of the decomposed results from the univariate ITD On the other hand,the direct multivariate ITD(DMITD)realizes the multivariate extension by developing key concepts including the multivariate extremum,multivariate baseline-node and baseline operator.Despite that they achieve the decomposition in different ways,both of IMITD and DMITD are naturally extended from the definition of the univariate ITD· Testing nonstationarity from industrial control loops is of great importance since it directly affects the accuracy of existing methods for oscillation diagnosis.Under the framework of hypothesis test,a new statistic based on the auto-covariance function(ACF)is proposed Then by analyzing the statistical difference between the process data and its surrogates,the potential nonstationarity can be easily detected.Secondly,to diagnose the poor control loop performance,a nonlinearity detection method that combining the higher-order statis-tic(HOS)and the time-frequency surrogate is developed.This work first presents a new cepstral definition of bicoherence,namely,bihocerence,to build the nonlinearity statistic.Then the LMD based de-trending and re-trending procedures combined with the random phase surrogates are utilized to determine the confidence limitAiming at promoting the application of time-frequency tools for industrial oscillation detec-tion and diagnosis,this work has studied and developed a number of new decomposition methods.Since they have provided new foundation for time-frequency analysis of multivariate data,more widespread applications like the signal denoising,biomedical engineering,and image fusion can be expected.