Ghannouchi, FadhelRawat, Meenakshi2017-12-182017-12-182012Rawat, M. (2012). Artificial neural networks for modeling and digital predistortion for software defined transmitters (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/4951http://hdl.handle.net/1880/105952Bibliography: p. 139-151Some pages are in colour.Includes copy of copyright permissions. Original copies with original Partial Copyright Licence.The overall objective of this thesis is to develop and analyze efficient and robust artificial neural network methodologies and distiibuted structures for complete transmitter modeling and its practical use as digital compensation solution for nonlinearity and hardware impainnents in wireless transmitters for software defined radio Applications. A suitable feedforward topology namely real valued focused time delay neural network is proposed and various nonlinear optimization algorithms are implemented to achieve best perfonnance in the presence of different power amplifiers and signals. While conventional digital predistortion (DPD) techniques focus mostly on power amplifiers and are dependent on signal statistics, the proposed linearization is more robust to signal statistics and generic in the sense that it adapts to any change in the input data even in the presence of modulator gain/ phase imbalances and DC offsets. Although highly robust, back propagation based feedforward neural network solutions have shortcomings such as high number of parameters to be stored leading to higher digital processing cost. Therefore, as an alternative cost cutting solution, this thesis ventures to modify conventional memory polynomial by applying layered structure similar to neural networks. With experimental results of different PAs, it is established that proposed three-layered-biased-memory-polynomial model enjoys better numerical stability and lower dispersion of coefficients which eventually helps in decreasing processing load on DSP and therefore can replace conventional memory polynomials providing similar performance. Above stated DPD techniques works in batch mode and when PA characteristics cannot be assumed constant over long time and constant adaptation of coefficients is needed, they may still lead to higher processing time therefore thesis further analyzes spatially distributed or lattice neural networks for adaptive digital compensation. It is reported that with its spatially distributed structure total processing cost is even lower than previously reported conventional adaptive nonlinear filters with reasonable performance especially in case of highly nonlinear PAs.xvii, 159 leaves : ill. ; 30 cm.engUniversity of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.doctoral thesis10.11575/PRISM/4951