Efficient and automatized error pattern modelling with hidden Markov models in digital communication

In digital communications error pattern modelling based on hidden Markov models is widely used to describe and generate error sequences. The length of measured error patterns are in the order of 105 and 106 to guarantee stochastic expressiveness, which results in high calculation times to parameterize the models. To reduce these computational efforts on the one hand and to free the user from the training period for HMMs on the other hand, this paper introduces an efficient and automated architecture for modelling binary error patterns. All model parameters are calculated transparently from the measured or simulated error pattern without any user interaction. Efficient parametrization is achieved by performing several optimization techniques such as state splitting, k-means based initialization, scaling reduction, equivalent diagonal models, and compiler optimizations. The architecture is firstly introduced for continuous error distributions which come true if the transmitted entities are processed on their own. In case of group-wise processing like in UMTS block transmissions, the error distribution can exhibit steps which is demonstrated by a measurement in a real UMTS test environment. In order to reproduce similar statistics an extended model is described which is also covered by the mentioned architecture.