Adaptive neural network control with inverse model: Application to the control of turning process

This dissertation presents adaptive on-line neural network control schemes for unknown dynamical processes and eventually for the control of feed force in turning process. The research consists of four tasks: Development of a machining process model, development of a neural network control structure and adaptation schemes, machining experiments to collect data for building a simulation model, and application of proposed control schemes to turning process.; By studying the process models in the literature, a turning process model on a CNC lathe is developed. The proposed model is nonlinear and sampling time is independent of spindle rpm. The characteristics of noise and the parameters of cutting dynamics are determined by the experimental data. To collect cutting force data, a SAE 6150 steel is turned with 25 sets of machining conditions on a 20HP LeBlond 1610 lathe.; The proposed adaptive inverse model (AIM) control scheme has similar structure as the model reference adaptive controller (MRAC). In AIM control, neural networks are used as the system identifier and the controller. The system identifier and the controller are first trained off-line to be a forward model and an inverse model of the process respectively. The off-line training is simultaneously done for the same set of data with different arrangements. During on-line adaptation, the system identifier is iteratively used to convert tracking errors into control errors and to calculate the current and past tracking errors after changing the weights of the controller. To be robust to noises, the controller not only considers the current error but also looks back to the past to minimize the "would-have-been" tracking errors since the change in weights of the controller would have generated different control inputs in the past, and, in turn, different tracking error.; Two implementations of AIM controllers, using feedforward networks and using cerebellar model articulation controllers (CMACs), are applied to control the feed force in turning along with an MRAC for comparison. Simulation results show that AIM controllers, which are designed without structural information about the dynamics of the process, perform comparably well with the MRAC. Especially for the process with severe noise, AIM controller with feedforward networks outperformed the others.; This research also provides a collective view on neural networks in on-line process control. Real-time back-propagation (RTB) and back-propagation through time (BTT) learning algorithms are presented for multilayer neural networks with internal time-delays. Convergent CMAC learning algorithms for on-line process identification are developed with proofs.; This research is unique in that we can design a control system without knowing the dynamics of the process. The proposed control system can outperform other control techniques since it is based on neural networks, which are noise-tolerant and can accurately represent nonlinearity. Application of neural network-based control schemes to the machining process is also notable. The parallel structure of neural networks can make integration easier with other concepts such as tool-wear monitoring, intelligent control, and supervisory control.