Blind identification and deconvolution filter design

Blind identification and deconvolution filter design are the two major topics in this dissertation. The system addressed in this dissertation is a two-input-one-output system, but only one input acts at a time. Because one of the inputs is known and the other is unknown, the objective is to identify the transfer function between the unknown input and the output. The denominator of the transfer function can be determined by an input-output model identified from the known input and output. If the system is stable, the numerator coefficients of the transfer function can be estimated by blind identification.; When the system is driven by an unknown input and the output is oversampled, the separate output sequences can be obtained by rearranging the sampled output. If the system is stable, a relation between these two output sequences can be used to identify the numerator coefficients of the transfer function. If the unknown input is a continuous low-frequency signal, it is necessary to approximate the continuous input by a discrete signal to use blind identification. For sufficiently low-frequency continuous inputs, blind identification can estimate the parameters accurately. If the input signal is a discrete sequence, a similar relation between the two oversampled output sequences can be constructed. Then, the parameters can be identified by blind identification without the low-frequency limitation on the input sequence. However, the estimated parameters are very sensitive to noise whether the unknown input is continuous or discrete.; Both finite-impulse-response (FIR) and Kalman filters are proposed in this thesis to restore the unknown input signal. For the FIR deconvolution filter design, a training signal is required during the design procedure. A shaping filter is introduced to design the deconvolution filter when a known frequency band of input is of special interest. Unlike the FIR deconvolution filter design, Kalman filter designed is based on the known estimated system dynamics without using a training signal. If the shaping filter is properly designed, a Kalman deconvolution filter can reconstruct the desired frequency band of the input signal. Experimental results are shown for the blind identification and the deconvolution filter design. Finally, conclusions and further research are discussed.