SPHT-LSTM: a novel industrial equipment condition prediction method

Abstract Production safety and efficiency depends on the reliability of the critical components in industrial equipment. However, signals measured by sensors may have complicated nonlinear and long-term sequential dynamics, which are difficult to predict faults correctly. To overcome these challenges, especially the detection of the subtle early-stage fault features and the ultra-long faults modeling, in this paper, a new SPHT-LSTM model based on hierarchical Transformers and LSTMs have been proposed. Multi-scale feature decoupling and dynamic fusion is made possible by the dual-channel architecture (superposition S1 long-term trends and progressive P2 short-term variations) based on the temporal sensitivity of LSTMs and the global dependency capturing of Transformers. Also, a Mean Performance Degradation (MPD) measure is proposed to measure degradation using segmented mean analysis to minimize noise and increase saliency of fault features. The univariate/ multivariate dataset and engineering data experimental validation reveals that SPHT-LSTM attains a prediction accuracy of 89% and 96% of early weak faults and accelerated degradation stages, respectively, which is much higher than that of the traditional techniques. Such findings indicate its strength and usefulness in condition-based maintenance (CBM) in the industrial environment.