Neural network control of a rotating elastic manipulator (Q5949147)

The authors discuss an important problem in robotics that has an analogy in aerospace applications, for example the dynamics of a helicopter rotor. However, this is not an analysis of dynamics of a rotating beam subject to torque applied to its hub, which is the control. This whole system is treated as a learning process analogous to a neural network methodology. The equations of state derived by the authors are frightening. And even then they have omitted some terms that may be important in their applications, such as Timoshenko's correction due to the inertia of the rotation of the cross-sectional area of the beam, and possible large deflections which would really mess up the equations of state. For higher speed rotors one should really consider the nonlinear relations between the appropriate stress and strain tensors. People with lots of energy and masochistic tendencies may consider the full set of constituent equations in the 1963 classic book of Novozhilov on nonlinear elasticity. The strange part of this discussion is that the derivation of the Hamiltonian of the motion would have more ``frightening'' terms, but can be computed in an identical manner to the one used by the authors to derive their equations of motion. Of course the authors never intended to use the system they displayed. They quickly retreated, and linearized everything, obtaining the Euler-Bernoulli model with bending moment at \(x=0\) equal to the applied torque \(T(t)\) plus the d'Alambert inertia term. Now when the system is linear, the Laplace transform \((t\to s)\) is applied. A fourth-order ordinary differential equation with respect to \(x\) and a single algebraic equation involving the control \(T(s)\), with four boundary conditions, is derived. An elementary solution \(U(x,s)\) of the fourth-order system is the usual combination of sines, cosines, sinh and cosh terms. The transfer function: input\(\to\)output (ratio of angle of rotation to the applied torque \(T(s))\) is represented as a ratio of two infinite series in \(s\). Obtaining an exact solution is an impossible problem. However much is known about this type of mechanical system. For a proportional controller, the root locus technique reveals that all closed loop poles and zeros alternate on the imaginary axis. Presence of the poles on the imaginary axis implies marginal stability. Thus a proportional and derivative controllers are proposed, moving all closed loop poles into the left half side of the complex plane. However the design of the controller for this system is still to be determined. The authors now regard it as a neural system so that a nonlinear regression algorithm could take the form of a learning process. The authors give a sketch of a multi-layer feedback design, where the training network guarantees improvement in stability. Error signals are sent back, and the last output layer affects the input (the torque applied). The article ends with a description of simulation, involving sketches of the back propagation neural architecture. The authors claim that their model is very effective and shows good robustness to perturbations.