While W. C. Mead was on Sabbatical at the Center for Non-Linear Studies (CNLS, 1991), he collaborated with other LANL staff members, technicians, and students to complete a Negative Ion Source Controller project that had been in progress for about 18 months. In the final six months of the project, he wrote a Nonlinear Adaptive Controller, called CTL, that used the CNLS-Net as a self-adapting model of the ion source. This work has generated considerable interest and is, to date, Dr. Mead's most widely published team research project. Details have been discussed at several technical meetings, published in the physics literature and, most recently, in the Handbook of Neural Computation (Oxford University Press, 1996). The work will be summarized briefly here. Readers wishing further detail will find it in the published reports on the project.
This project was carried out on the LANL Discharge Test Stand's Small-Angle Negative Ion Source, a high-brightness source of negative hydrogen ions being developed for space-based accelerator use. The source is a small volume, high power, pulsed device with complicated internal plasma physics and chemistry. Although its operation is understood in general terms, details of the source's behavior are unpredictable and not well-understood.
The figure below shows the Negative Ion Source from a control viewpoint: a "black-box" that has inputs and outputs. The four main inputs that can be set by a controller are the Anode Temperature (Ta), the Cathode Temperature (Tc), the Arc Supply Voltage (Vas), and the Hydrogen Gas Flow Rate (H2 Flow). When operated by a human, the computer that handles the low-level interface to the ion source presents a manually adjustable control console display. The low-level control computer makes measurements while the source is in operation, and these are the "outputs" from the ion source. Four scalar readouts monitor the actual values of the four quantities that were set as control inputs. PID feedback loops automatically adjust the machine inputs so these readbacks match the controller-set values. In addition, the diagnostics provide three waveforms that monitor the electrical operation of the pulsed ion source: the Arc Voltage (Varc), the Arc Current (Iarc), and the Beam Current (Ibeam). The high-level controller (for example, the human operator) seeks to adjust the four input/control settings so that the output of the ion source has "Good Beam" characteristics. The characteristics that define "Good Beam" are
Prior to this work, the Negative Ion Source had been successfully controlled by trained human operators. However, two previous attempts at automatic high-level control had failed. Several performance characteristics of the source led to the severe difficulty in achieving the required high-level optimization and control.
Dr. Mead wrote the CTL Nonlinear Adaptive Controller to meet these substantial challenges. A schematic representation of CTL is shown in the drawing below.
Much of the apparatus of CTL is devoted towards data handling and transformation, in order to present the information to CNLS-Net in the form the Neural Network requires. For example, the Input Processor converts the diagnostic beam current waveforms into a "figure of merit" (fom) that expresses the "Good Beam" criteria above in a monotonic scale of 0.-1.0. This process is illustrated in the sample beam current pulses shown in the figure below (NOTE: The beam current is plotted below on an arbitrary scale; the desired current pulse, seen in the Wv 577 frame, is actually negative-polarity). During the collection of these waveforms, the Negative Ion Source was functioning in various of its several operating modes. Some modes show parasitic oscillations, other kinds of noise, or adverse intra-pulse or inter-pulse dynamic behavior. Waveform 577 is the "near-ideal" behavior with a strong negative current pulse during the entire "extraction window" (indicated by the vertical bracketed region of the pulse). Waveform 165 is nearly as good; however, the pulses preceding and following the pulse shown were unsatisfactory, so the fom is reduced by the pulse-to-pulse non-repeatability. The upper three waveforms are obviously flawed by noise and/or by reduced average beam current.
Having now established a viable "figure of merit," the task is ready for CNLS-Net. The NN's function is to learn how the machine operates, construct a model of that operation from sparse data in 4 dimensions (the independent control variables), and advise the CTL controller how to set the control parameters in order to maximize the fom. It's a non-trivial task, but the NN does a good job. The results of the CTL-optimized tunings are shown in the waveforms below, which were taken from 6 independent runs over several hours of Negative Ion Source operation.
The success of the CTL optimization can be seen by comparing these six optimized waveforms, which repeat to within about 5-10% in beam current. Optimization and control by the NN-based adaptive controller compared favorably with that achievable by trained human operators.
This project was supported by the DOE and was performed as a Laboratory Directed Research and Development project.
You may use your browser's 'Back' button to return to the most recently viewed ANSR page.
Return to ANSR Home Page or Site Contents -- or -- Contact ANSR.