Masterarbeit - Details
Active Noise Cancellation Using Nonlinear Adaptive Filters and Machine Learning
Supervisor: Till Hardenbicker
Area: Kernel Adaptive Filters, RBF Networks, Active Noise Cancelling
Tools: Matlab, C or Python
Apart from its well known use in consumer headphones, Active Noise Cancellation (ANC) is valuable for hearing protection at extreme sound pressure levels. Possible fields of application involve manufacturing halls, airports or race tracks.
ANC aims to block environmental sound by generating a compensation signal with a complementary phase. In most cases, this signal is obtained by filtering the recorded environmental signal with an adaptive filter. We may deploy the Normalized Least Mean Square (NLMS) algorithm, to approximate an optimal filter for the involved acoustic paths. This algorithm is a simple but efficient estimator for linear systems, however it fails to learn nonlinear system properties. In ANC, nonlinearities can occur when the internal speaker is dealing with signals of high amplitudes as in the situations given above.
Luckily, the NLMS algorithm’s inability to map nonlinearities can be overcome by expansion to the Kernel-LMS (KLMS) algorithm. By mapping the input data onto a high dimensional feature space, it is able to learn nonlinear relationships. Opposing to its linear variant, the KLMS algorithm requires a dictionary of representative input vectors that is updated online.
The task of this thesis is to expand an existing NLMS based ANC topology by the KLMS algorithm. Different update rules for the employed codebook shall be investigated, regarding their performance and complexity. Since they have a close relationship to the KLMS algorithm, Radial Basis Function (RBF) Networks should be investigated as well. The analysis should reflect reasonable constraints for the computation in real time and can include real world example signals or even real time measurements.