Multiple-prior-knowledge neural network for industrial processes

A novel multiple-prior-knowledge neural network for industrial processes is proposed. Diversity of priori knowledge from industrial processes are discovered and embedded into three-layer hybrid feedforward neural network in the form of penalty function or nonlinear constraints. Meanwhile, Sequential quadratic programming method is selected as the network learning algorithm. Compared with BP neural network and an industrial neural network (the Bounded Derivative neural network, BDNN), MPKN network not only satisfies a certain process mechanism and keeps the process safety, but also has better generalization. Its effective performance is validated with the modelling of a nonlinear process of Continuous-stirred tank reactor.