Atsushi Aoyama

[Show abstract] [Hide abstract]

**ABSTRACT:** Nonlinearity is the rule rather than the exception in chemical processes. Neural networks are considered to be attractive for modeling nonlinear processes because of their ability to approximate arbitrary functions. However, previous attempts to use neural networks in the internal model control framework were not very successful as the inverse model generated by the neural network was not very accurate and lead to performance degradation. To overcome this problem, the use of steady state data in addition to the transient data for the network training is proposed. In general, neural network models are empirical nonlinear input-output models, with all the input-output mapping details hidden in the structure and the weights of the network. However, in the present research, a specific model structure consisting of the bias/drift term and the steady state gain is imposed, leading to the possibility for greater model insight and facilitating an analytical model inversion. First, a method to integrate mathematical models with neural networks, where a mathematical model models the bias/drift term and a neural network models the steady state gain, is proposed. This approach is applied to a SISO control-affine process control and showed comparable performances to the exact mathematical based IMC. Second, a modeling scheme where both the bias/drift term and the steady state gain are modeled by neural networks is proposed. The proposed approach is applied to the tasks of modeling and control of SISO nonlinear processes which are not inherently control-affine and showed a significant performance improvement over a conventional PID controller. In addition, this control-affine model based approach was combined with the generalized Smith predictor approach for the control of nonminimum-phase nonlinear processes. When the process is poorly described by a control-affine model, two approaches are proposed. First, a third neural network is introduced to model the discrepancy between control-affine model and actual process and incorporated in an auxiliary loop. Second, a modeling scheme using a fuzzy neural network structure is introduced. Both approaches lead to improved closed-loop performance. ETD Collection for Purdue University.