A Fault Detection Model Based on Ensemble Slow Feature Analysis

To tackle the problem that single slow feature analysis algorithms encounter difficulties in fault detection owing to the coexistence of multiple industrial characteristics in continuous industrial processes, an ensemble slow feature analysis (ESFA) algorithm for industrial process fault detection is proposed. Initially, exploratory data analysis is carried out on the industrial process data to preliminarily identify theoretically suitable preset detection models. Subsequently, these models are optimized by means of a genetic algorithm to select appropriate process variables and generate fault detection results. Next, a model selection algorithm is utilized to identify the most appropriate preset detection models as sub-models for the ESFA. A model result fusion algorithm is devised, adopting a Bayesian fusion strategy to combine the statistics of the sub-models so as to generate new detection values and control limits. Finally, the performance of the ESFA is validated through the Tennessee-Eastman (TE) industrial process case study. Experimental results show that ESFA can effectively select sub-models and integrate their advantages, resulting in enhanced fault detection performance.