Multivariate statistical modelling and monitoring for nonstationary industrial process with slow feature analysis and temporal working pattern partition

Due to change of working condition, nonstationarity widely exists in modern industry, leading to time-variant statistical properties and frustrating the modelling techniques for stationary process. In this research, slow feature analysis and working pattern partition based method is proposed for nonstationary industrial process modelling. Slow feature analysis is employed as a filter to separate the signals with different frequencies. Next, stationarity test is applied to the ordered slow features. Then, four categories of features can be obtained with specific significance. Among them, the slow nonstationary features represent the inner change of working condition and thus can be adopted to partition the working condition into multiple patterns, temporally. Hence, the nonstationarity can be decomposed into piece-wise stationarity in the feature space. Each working pattern can then be modelled separately, using the normal data based on stationary techniques. Finally, the proposed method is validated on a real copper flotation process.