Estimating Angular Diffuse Components and Model Misspecification in MIMO Channel Sounding

Parameter estimation for multiple-input-multiple-output (MIMO) channel sounding data aims at accurately describing channel measurements with physically realistic and interpretable parameters. The performance of model-based approaches, e.g., maximum likelihood, is determined by the accuracy of the imposed signal model. For channel sounding data, it has turned out to be beneficial to use two distinct concepts for the description of the propagation process. The specular components account for the dominant propagation paths of plane waves, whereas diffuse components model the weaker but more diverse propagation processes by means of a colored noise process. In order to improve the accuracy of the model for the diffuse components, we propose a simple but still flexible parametric covariance model that allows us to account for a smooth power angle profile that describes the correlation in the spatial domain. Moreover, the model for the deterministic part of the signal is usually contaminated by calibration errors, which in turn deteriorate the reliability of the specular path estimates. This is most prominently visible by the estimation of so-called ghost paths. To mitigate this, we introduce a new model order selection scheme based on the so-called misspecified Cram & eacute;r-Rao bound (MCRB) that accounts for the unavoidable modeling errors. In addition, to avoid the fitting of ghost paths caused by the faulty modeling of strong specular components, we locally decrease the estimated signal-to-noise ratio (SNR) in the time domain around already estimated ones. Furthermore, as these changes to the signal model require more computational resources compared with existing algorithms, we also showcase how necessary quantities like likelihoods, score functions, and Fisher information matrices can still be computed efficiently. We implement our proposed extensions within the Richter maximization approach (RIMAX) framework. We also showcase that they improve the reliability of the produced estimates compared with plain vanilla RIMAX on real measurement data.