Multivariate Feature Extraction Based On Residuals And Its Application In Process Monitoring

Maintaining a process operated at steady state is the goal of continuous operation in chemical industry,which requires process faults timely detected.In previous studies,the idea of fault detection was to detect the difference between the current process and normal operating conditions.Under an ideal operating condition,the deviation can be approximated by the residuals of real measurements relative to mathematical models if process mechanic model is available.However,as the information available in the actual process is limited,such as stream composition,equipment status and process mechanism,the process information cannot be accurately extracted by mathematical model.The residuals of a model cannot be used as a criterion to identify the deviation of a process,let alone to recognize a process failure at its early stage.With well-developed distribution control system(DCS)in process industry,large amount of data have been collected,which have greatly accelerated the development of data-driven process monitoring methods.Under a specific normal operating condition,correlation among variables are relatively stable.The system can be well approximated by a data-driven multiple regression model,with a modelling residuals.Based on this,a new residuals-based process monitoring method is proposed.Based on the topology of a process,variables are selected to establish a regression model to extract the correlation for a steady state operation,especially those with spatial correlations regarding certain unit or unit group.Then,the principal component analysis method is applied to extract the main information of residuals.By monitoring main feature change in prediction residuals,the early identification of the process fault is realized,and the influence of noise on the process monitoring is avoided.Data from the benchmark Tennessee Eastman process and a pre-reforming reactor are researched to validate the proposed methods.The results show that process feature,especially the spatial correlations among variables,can be accurately extracted by regression model.Compared with the traditional Principal Component Analysis algorithm,the process monitoring model based on residuals can detect TE process faults earlier.In the case of pre-reforming reactor,catalyst sulfur poisoning fault can be detected by proposed method at least 13 hour earlier than human operator.