Vibration Control Update

by Gordon Kraft

In the last edition of Signals and Noise we reported that the ECE Robotics Laboratory received a grant of $365,000 from the National Science Foundation for research related to the control of unwanted vibrations using CMAC (Cerebellar Model Articulation Controller) neural networks. We are now in the second year of this research project and there are some promising results.

Most vibration controllers are passive systems like the suspension and rubber mounts in one's car. When better performance is required, active feedback systems are used. Active systems measure the vibration using sensors, such as accelerometers, and act to produce a force to oppose the vibration. In our research the active part of the controller is augmented with a "learning" system called CMAC. The CMAC system is a form of neural network that remembers what the overall system does. That is, it learns through practice what control signal is appropriate during the next cycle. The advantage of using CMAC is that the system’s performance continues to improve as it operates. This allows the controller to "tune" itself as it operates. For example, if the disturbance frequency changes, the network can adjust and compensate for the frequency shift.

To investigate the effectiveness of the CMAC controllers, several laboratory experiments and computer simulations have been performed during the past year. Jeremy Pallotta (MS  ‘99) developed a new version of CMAC with "weight smoothing" which is more appropriate for vibration control. Jeremy produced a digital simulation in MATLAB that verified the control algorithm. This concept proved very effective for a single-input single-output second-order model with a single frequency disturbance. Simulation results led directly to a real-time audio range canceling system. The system consisted of an audio speaker (the vibration disturbance) and a microphone to sense the unwanted noise. A sinusoidal tone at 100 Hz was used to drive the speaker. The output of the microphone was sampled at 6500 Hz by a computer and used as the training signal for the CMAC network. The network learned quickly to produce exactly the correct signal to cancel the unwanted tone. This system was tested under a variety of operating conditions, including multiple disturbance frequencies. The effects of learning rate, generalization, and feedback gain were experimentally studied. This was the first successful real-time vibration controller using a CMAC neural network. In parallel with Jeremy’s work, Jon Frain (MS ‘99) used the traditional CMAC algorithm to approach the same problem. A comparison was made between the effectiveness of the traditional CMAC controller and the enhanced "weight smoothing" CMAC controller. Jon’s audio frequency vibration canceling system was also very effective at reducing the unwanted vibrations. These results led to two papers presented at the American Control Conference in San Diego (June 1999).

Also in parallel with the real-time studies, Yongjun "Jim" Wu (MS ‘99) developed a method to incorporate CMAC into a program called Working Model from Knowledge Revolution Company. This program allows a user to visualize the operating vibration controller on a computer screen. Jim’s program simulated a beam suspended on springs subject to a sinusoidal vibration at one end. Acceleration measurements at the other end of the beam are used as the training signal for CMAC. The network generates a control signal that can drive an actuator to cancel the vibration source. Jim’s program also allows a user to visualize the CMAC weight space as the controller is actively learning. These results will be presented at the ACTIVE conference in Florida (December 1999).

Work is now continuing in the ECE Robotics laboratory to extend these ideas to higher order systems. This summer a team of people including faculty members Dr. Sivaprasad and Dr. Andy Kun; master’s degree students Peter Bick, Paul Wheeler, Kangyong Liu, and Rob Smith; and undergraduates Mark Sinclair, Bill Roy, Boris Goranovic and Cam Trinh are all working on vibration control problems. Hopefully next year we will be reporting their results.