Emulating human process control functions with neural networks

This investigation demonstrates that neural networks can perform some of the tasks in controlling complex systems that have been traditionally reserved for humans. Neural networks can be used to fuse different types of knowledge from many sources into a general process model. This technique allows process models to be formed for systems that are too complex to be modeled with conventional tools. By adding relatively few local measurements, a general process model can be calibrated into a numerically accurate local model of the process. This local model can then used for steady-state process optimization. The architectures and training techniques needed to produce neural networks capable of performing these functions are discussed. This technology was applied to the control of a complex system--a grain harvesting combine. Field tests of the harvesting process under neural network control demonstrated that the controller was robust and capable of exceeding the performance of expert human operators.