Identifying and Modelling Multipath Clusters in Propagation Measurement Data Ghassan S. Dahman [1] , Robert J. C. Bultitude [2] , and Roshdy H.M. Hafez [1] [1] Department of Systems and Computer Engineering, Carleton University, Ottawa, Canada [2] Communications Research Centre, Ottawa, Canada {gsdahman, hafez}@sce.carleton.ca, and robert.bultitude@crc.ca Abstract- In this paper, a new algorithm that is able to identify and track multipath clusters in the delay-angular domain is introduced. The introduced algorithm avoids calculations of a single distance measure from quantities with different nature; instead it performs the clustering via extracting multiple 1D waveforms from the 2D delay-angular domain. The results from applying this algorithm to measured radio propagation data (recorded using a sounding system with 32 receive antenna) is used to classify the identified multipath clusters into different groups. The lengths of the active intervals (which are defined as the lengths of the intervals in which clusters exists) of the multipath clusters as well as their powers in the different groups are modelled. It was shown that the power of multipath clusters in these different groups can be modelled using the Generalized Extreme Value distribution and that this model passes the Kolmogorov-Smirnov (KS) goodness of fit test at the 5% significance level. Keywords: multi-antenna channel modelling; multipath clustering; sequential-delay-angular clustering. INTRODUCTION Results from radio propagation measurements using mutliple-antenna sounders have shown that multipath components (MPCs) do not arrive uniformly in delay-angular space. Instead, energy is concentrated in clusters linked to reflections, scattering, and/or diffractions caused by interacting objects [1], [2]. Multipath clusters have been defined in the literature as accumulations of MPCs with similar delay-angular parameters (e.g., angle-of-arrival (AOA) and delay) that share the same long-term evolution i.e., they stay intact over time [1], [3], [4], [5]. Different Multiple-Input Multiple-Output channel models have employed the concept of multipath clusters intensively. However, identifying and tracking the time-varying characteristics of multipath clusters is still an open topic for research that is addressed by different researchers using different approaches [4], [6]. Visual inspection was originally used to identify multipath clusters. However, this is a time consuming process especially for large data pools [1], [5], [7]. To make the identification of multipath clusters more practical, different automatic clustering algorithms have been introduced to group MPCs into clusters in delay-angular space. The different clustering algorithms can be categorized into two main groups: 1. Joint-Delay-Angular clustering: When these approaches are used MPCs are grouped into clusters using a total distance measure that describes how certain MPCs are close to (farther from) each other jointly in both the delay domain and the angular domain. In accordance with these algorithms, the calculation of a single distance measure extracted from different quantities (e.g., delay, and AOA) represents a major challenge due to the different nature, and hence the different units, associated with the quantities involved. Two solutions have been suggested to solve this challenge: the use of parameter normalization [6], and the use of the Kernel density estimation technique [4], [8]. However, tracking the behaviour of clusters with time using such algorithms requires the identification of the clusters within each snapshot, then pairing the different clusters among different snapshots. 2. Sequential-Delay-Angular clustering: With this approach, the clustering is performed first in the delay domain and then in the angular domain conditioned on the corresponding cluster delay [9], [10]. The sequential delay- angular clustering approaches are attractive over the joint delay-angular approaches because of two reasons: 1) using the sequential delay-angular clustering approach simplifies the multidimensional clustering problem by transforming it into multiple sequential 1-D clustering problems, and 2) applying the sequential delay-angular clustering approaches allows for using meaningful measures in each different dimension (i.e., delay vs. angular), which eliminates the controversy arising from normalizing two quantities with different natures to come up with a universal normalized value that is needed for the multidimensional clustering algorithm. The contribution of this paper is two-fold. First, a new algorithm to identify multipath clusters from measured multiple-antenna radio propagation data is introduced. This algorithm is used to estimate the powers and the length of the active intervals, which are defined as the length of the intervals in which multipath clusters exist, of multipath clusters. Second, a new method is used to classify identified multipath clusters into different groups enabling the powers of the clusters in each group to easily be modelled. Section II gives specifications of the channel sounder and propagation measurement procedures. In Section III a new algorithm to identify and track multipath clusters is proposed. Finally, results are reported in Section IV. II. PROPAGATION EXPERIMENTS Measurements were conducted in downtown Ottawa, a typical North American urban area. The Tx monopole was mounted at a height of 6 m above ground level on a mast that was extended from the roof of a small trailer parked at a fixed curb-side on a busy street. The Rx antenna was a switched uniform circular array (UCA) with 32 quarter-wavelength 978-1-4244-3574-6/10/$25.00 ©2010 IEEE