Statistical Process Monitoring Method Based On Slow Feature Analysis

In recent decades,with the improving level of large-scale industrial production,the gradually complex production mode and the increasingly high demands for products,the safety performance of the production procedure becomes more and more significant.Therefore,it is necessary to establish an efficient and stable monitoring model to monitor each production link,ensure product quality and reduce safety accidents.The data-driven dimensionality reduction methods can extract the essential information of system from the noisy signals and establish the monitoring model to evaluate the process performance.This dissertation treats Slow Feature Analysis(SFA)as the core algorithm,and takes researches and improvements to SFA on the consideration of outliers,system dynamic properties and the relations between process variations and output in industrial data,respectively.(1)To address the deficiencies of non-robustness of the conventional SFA method,this section proposes a robust SFA monitoring modeling method.It adopts robust processing to SFA using an improved M-estimator.First,outliers are subtly identified by robust approaches in the eigenvalue decomposition step,and given less weight to reduce the degree of deviation from outliers to proper data distribution.Then,the proposed method iteratively eliminates the steady and dynamic influence caused by outliers separately,reconstructs normal data and establishes monitoring indicators.Through the comprehensive monitoring of the static and dynamic properties in the system,the understanding of the process states can be deepened.Finally,the effectiveness of the proposed approach is illustrated with a numerical nonlinear system and TE process.The results prove that the robust algorithm can build healthy monitoring model to properly distinguish normal changes of operation statuses from real system faults.(2)The varying speed of variables and serial correlation in the process are studied.And a comprehensive monitoring modeling approach based on serial correlated dynamic SFA is proposed.SFA focuses on the varying speed between nearby samples,whereas temporal correlations often exist in industrial processes.Owing to this two dynamic properties existed,this section adopts a new objective function to extract slow features which have serial correlations.Then,a comprehensive monitoring model is constructed to describe the dynamics of the process more properly.To verify the effectiveness of the proposed algorithm,TE process is employed.By sufficiently explore the dynamic characteristics in industrial process,the proposed monitoring strategy can provide more accurate results.(3)SFA extracts slow features in an unsupervised manner,which have low correlations with outputs.In this work,a complete monitoring strategy based on quality-correlated SFA method is proposed.This section introduces Canonical Correlation Analysis(CCA)approach,proposes a new algorithm named CCSFA which concurrently considers the output correlations and slowness.This strategy could extract the slow features which are correlated to outputs from process data.It can divide process data space into quality-relevant and process-relevant subspaces in supervised manner.Then,comprehensive monitoring models are separately built in subspaces to explore the characters of process performance and system output in different conditions.Finally,TE data is adopted.It turns out that the proposed model can provide a meaningful and convictive practical interpretation about process dynamics and quality variable due to the detailed analysis and monitoring of process data.