A novel adaptive extreme learning machine-autoencoder algorithm for online monitoring and fault detection

Abstract

The extreme learning machine-autoencoder (ELM-AE) algorithm has attracted significant attention with regards to the online monitoring and fault detection of industrial process in recent years. However, ELM-AE algorithm generally observes increased false alarm rate (FAR) when it is applied to complex industrial process monitoring with time-varying characteristics. To solve this problem, a novel adaptive extreme learning machine-autoencoder (AELM-AE) algorithm for industrial process monitoring is proposed in this paper. The AELM-AE is implemented by embedding the approximate linear dependence (ALD) method into the conventional ELM-AE algorithm. The ALD condition is used to select new and more valuable online samples to dynamically update the previous monitoring model, which enables the proposed AELM-AE algorithm to effectively adapt to time-varying industrial process monitoring. Through extensive experiments on the benchmark Tennessee-Eastman process, it is demonstrated that the newly proposed AELM-AE algorithm can improve the fault detection performance and significantly reduce the FAR compared with the conventional ELM-AE.