| Piezoelectric ultrasonic motor (USM) is a newly developed motor in 1980's. It can be derived directly and has many useful features such as high torque at low speed, silence, compactness in size and no electromagnetic interferences. So USM has been used in many practical applications. The working principle of the USM is based on ultrasonic vibration force of the piezoelectric elements and mechanical frictional force. The operational characteristics of the USM are affected by many factors. Strongly nonlinear characteristics could be caused by the increase of temperature, the changes of load, driving frequency and voltage and many other factors. Thus the dynamic models of the USM are very complicated, which make it difficult for the traditional control method to fulfill the effective control of the USM. Recently the artificial intelligent methods based on neural networks and fuzzy systems have become the main approaches to perform USM control. Using these intelligent control methods, one has achieved good control effectiveness for various kinds of ultrasonic motors. However, the existing intelligent methods for the USM control have some shortcomings, such as complex structures, slower convergent speeds and lower convergent precision, and some others. Solving these problems is the key to improve the control efficiency of the USM. So it is very important to carry out further research on the control method of the USM using intelligent control methods and computer technique combined with the existing control methods. Aiming at the mentioned theoretical problems above, this thesis proposes a novel dynamical time delay neural network (DTDNN) for ultrasonic motors identification. The major contents are summarized as follow: (1) The characteristics of the USM, the comparisons of the control strategy, and the practical applications are reviewed. (2) The structure of the time-delay neural network is described and the on-line learning algorithms of the network are deduced. Furthermore, the system identification and control using neural networks are introduced. (3) A newly developed time-delay recurrent neural network identifier of USM is constructed. A bimodal neural network controller is designed where both the driving frequency and amplitude of the applied voltage are used as network inputs. Multilayer static networks transform the dynamic time-molding problem into a static space mold problem, which pass to change into the dynamic system linear or nonlinear discrete system. But the dynamic recurrent networks themselves import dynamic quantum in order to complete the identification of the dynamic system. In chapter 3, a three-layer DTDNN, which has two inputs and single output, is adopted to implement the identification for USM. The nonlinear activation function of the hidden-layer is the hyperbolic tangential function, the input and the output layer's is linear function. A dynamic recurrent back propagation algorithm is developed. To guarantee the convergence and faster learning of the proposed neural network model, the optimal adaptive learning rates are derived in the sense of discrete-type Lyapunov stability. Numerical results show that the proposed DTDNN identifier can approximate the nonlinear input-output mappings of the ultrasonic motor quite well. (4) A newly developed time-delay recurrent neural network controller of USM is proposed. The neural network is a kind of information processing system, which is used to imitate the structures of the human brain nerves, thinking and judgment etc. It can constitute the highly nonlinear dynamics system, and has some characteristics, such as the ability of large-scale parallel processing, self-adaptive, self-organize, self-study, the distribute storage, and so on. So the control system using the neural network method has the stronger adaptive ability and is more robust. In chapter 4, a three-layer DTDNN, which has two inputs and two...
|
PDF Full Text Request