Soft sensing based on artificial neural network

Soft sensing or inferential estimation has long been considered a potent tool to deal with the conflict between small control interval and large sampling interval existing in a wide variety of industrial processes. To extend the soft sensing from linear system to nonlinear case, we propose a nonlinear soft sensor on the basis of multi-step prediction using a recurrent neural network and a novel alternating training method especially suitable for slowly sampled primary output. The nonlinear soft sensor has been demonstrated by simulation results to be able to produce qualified estimation with good convergence speed