Towards self-learning sensors: Matching pursuit with dictionary learning on FPGA

Abstract: A general approach for sparse signal decomposition is matching pursuit with an over-complete dictionary of features. It has been demonstrated that this gives efficient codes when applied to images and acoustic signals. This thesis implements an on-line version of matching pursuit with Hebbian dictionary learning, which is theoretically able to process vibration and acoustic emission signals for condition monitoring of, for example, bearings. The implementation is done in C and critical parts are identified and implemented in VHDL with the goal of synthesising those parts on an FPGA. The C implementation can handle data in two formats, both fixed and floating point and is verified to be functioning in both cases. The VHDL components are verified in simulation. The maximum processing speed possible for the designed system is expected be on the order of one million samples per second, 1 Msps.