Adaptive dynamic inferential analytic stationary subspace analysis: A novel method for fault detection in blast furnace ironmaking process

Detecting faults in blast furnace ironmaking process (BFIP) remains a challenging task due to the hybrid properties involving dynamics and nonstationarity. To address this problem, this pa-per develops a novel method called adaptive dynamic interpretable analytic stationary subspace analysis (DiASSA). The method employs an inferential observation decomposition strategy to dis-tinguish dynamic, static, and nonstationary components from BFIP data. It then implements an iterative modeling algorithm to estimate dynamic consistent features within a closed region and effectively isolates the dynamics and statics. The static part is further modeled by ordinary an-alytic stationary subspace analysis (ASSA) to construct static consistent features and eliminate the interference of nonstationary information. Moreover, an adaptive fault detection strategy is developed, using exponentially weighted statistic structures and adaptive threshold settings to enhance detection efficiency and robustness. Theoretical investigations and case studies confirm the advantages of the proposed method over traditional methods.

Automated summary

This paper proposes a novel method called adaptive dynamic interpretable analytic stationary subspace analysis (DiASSA) for detecting faults in blast furnace ironmaking process (BFIP). The method distinguishes dynamic, static, and nonstationary components from BFIP data using an inferential observation decomposition strategy and estimates dynamic consistent features within a closed region through an iterative modeling algorithm to effectively isolate dynamics and statics. The static part is further modeled by ordinary analytic stationary subspace analysis (ASSA) to construct static consistent features and eliminate the interference of nonstationary information. An adaptive fault detection strategy is also developed, using exponentially weighted statistic structures and adaptive threshold settings to enhance detection efficiency and robustness. Theoretical investigations and case studies confirm the advantages of the proposed method over traditional methods.