The main goal of this work was to develop adaptive filter algorithms and test their performance in active combustion control. Several algorithms were incorporated, which are divided into gradient descent algorithms and pattern searches.

The algorithms were tested on three separate platforms. The first was an analog electronic simulator, which uses a second order acoustics model and a first order low pass filter to simulate the flame dynamics of an unstable tube combustor. The second was a flat flame, methane-air Rijke tube. The third can be considered a quasi-LDI liquid fuel combustor with a thermal output of approximately 30 kW.

Actuation included the use of an acoustic actuator for the Rijke tube and a proportional throttling valve for the liquid fuel rig. Proportional actuation, pulsed actuation, and subharmonic control were all investigated throughout this work.

The proportional actuation tests on the Rijke tube combustor have shown that, in general, the gradient descent algorithms outperformed the pattern search algorithms. Although, the pattern search algorithms were able to suppress the pressure signal to levels comparable to the gradient descent algorithms, the convergence time was lower for the gradient descent algorithms. The gradient algorithms were also superior in the presence of actuator authority limitations.

The pulsed actuation on the Rijke tube showed that the convergence time is decreased for this type of actuation. This is due to the fact that there is a fixed amplitude control signal and algorithms did not have to search for sufficient magnitude. It was shown that subharmonic control could be used in conjunction with the algorithms. Control was achieved at the second and third subharmonic, and control was maintained for much higher subharmonics.

The cost surface of the liquid fuel rig was obtained as the mean squared error of the combustor pressure as a function of the magnitude and phase of the controller. The adaptive algorithms were able to achieve some suppression of the pressure oscillations but did not converge to the optimal phase as shown in the cost surface. Simulations using the data from this cost surface were also performed. With the addition of a probing function, the algorithms were able to converge to a near-optimal condition.