| One typical problem in the design of linear particle beam accelerators is that of position and angle jitter. Jitter is the temporal variation of a particle beam parameter. In this case, the parameters of interest are the transverse beam position relative to central beam axis, and the beam trajectory angle. The current method used to correct these problems is closed-loop feedback control. Such systems have worked adequately in the past, but with the advent of new accelerator designs, they have started to fall short of the mark. It is proposed and shown here that many adaptive and predictive techniques can greatly extend both the bandwidth and the accuracy of current systems. These techniques include both adaptive feedforward and feedback methods. Included are both linear and nonlinear prediction systems, using both local and global prediction models. All of these techniques were developed for use on pulsed-particle beams. They can perform real-time correction of high-frequency jitter within the bounds of an individual pulse (intrapulse correction), or in the case of short pulse lengths, they can correct lower frequency jitter on a pulse-to-pulse basis (interpulse correction). The systems were designed and simulated on a digital computer. Actual position and angle jitter data were measured at Los Alamos National Laboratory, Argonne National Laboratory, and the Stanford Linear Collider. These data were used as input for the simulated control systems. For the intrapulse systems, a Kalman filter predictive method was able to reduce rms jitter by an average factor of four times over that of standard systems. The bandwidth of correction was increased by an average factor of 25. The Kalman method also worked well on the interpulse data increasing jitter reduction by a factor of three. An adaptive feedback method also gave excellent results with the interpulse data (reduction between two and three times that of standard systems) and can be easily integrated into existing feedback control systems. Other techniques included the autocorrelation method, the covariance method, the LMS algorithm, and various types of neural networks. With a few exceptions, results were outstanding for all the systems using all of the data sets. |
