A Novel Speed-Variable Gate Recurrent Unit with Nonlinear Exponential Transformation and Its Application to Soft Sensors

Gated recurrent unit (GRU), as a significant paradigm for data driven soft sensor modeling approach, is widely used in industrial processes for estimating key quality variables. However, the inherent linear regulation of GRU hidden states restricts efficient information transmission and thus limiting prediction performance. To address this issue, a novel speed-variable gate recurrent unit (SVGRU) network based on nonlinear exponential transformation is proposed. A speed-variable regulation gate is integrated into the GRU architecture to autonomously determine the speed adjustment factor at each time step, modulating the information flow rate in the gate controllers. The speed-variable regulation gate is combined with the update gate to perform a nonlinear exponential transformation, breaking the inherent linear constraint of SVGRU hidden states. This mechanism effectively enhances the information transfer rate of SVGRU, improving its information extraction capability. The case studies show that the proposed SVGRU outperforms existing methods in terms of convergence speed and prediction accuracy.