Application Research On Multivariate Empirical Mode Decomposition In Chemical Process Fault Diagnosis

For the current chemical process, the production technology is increasingly complex, and the size growing, coupled with the chemical data itself sharing the characteristics of high dimensional, large quantities, strong correlation and full of noise, etc., which all make production safety and product quality face greater challenges and higher requirements. In this paper, by using the capabilities of empirical mode decomposition (EMD) adaptively processing signal in accordance with the frequency scale to extract fault features, and the advantages of multivariate EMD (MEMD) processing associated information, their applications in chemical process fault diagnosis are researched, and combined with modified dynamic visibility graph (MDVG) algorithm, a data-driven fault diagnosis method is proposed, applied in Tennessee-Eastman (TE) process for real-time monitoring and online diagnosis, the simulation results verifying the effectiveness and superiority of the proposed method. The main research contents cover as follows:Firstly, with residual sensitivity to faults, a fault diagnosis method based on ensemble EMD (EEMD) residual is proposed. Six Sigma control charts are applied for historical data to define the control limits of the residuals fault diagnosis. Bayesian information criterion is adopted for online real-time data to determine the EEMD moving windows. The EEMD maximum residual changes are calculated based on moving window sampling data, and the time and cause of fault are diagnosed online by combing with the control limits of the residuals fault diagnosis.Secondly, in order to improve the efficiency and accuracy of the process monitoring and overcome the limitations of monitoring by single variable, propose a fault online monitoring method for multivariable process based on MDVG algorithm. The DVG algorithm is improved, which can detailed characterize networks properties of different time series data by data normalization and the introduced time interval constant, to minimize the mean of mode occurrence number of the concerned DVG properties.Finally, a fault diagnosis method is proposed by combining MEMD and MDVG. Determine monitoring variables by MEMD residual; determine monitoring indicators and thresholds by the historical data of various monitoring variables with MDVG and implement online monitoring; diagnose the fault under abnormal conditions cause by comparative analysis with MEMD residuals.