Adaptive learning techniques have recently been considered for structural health monitoring applications due to their flexibility and effectiveness in addressing real-world challenges such as variability in the monitoring of environmental and operating conditions. In this paper, an active learning data selection procedure is proposed that adaptively selects the most informative measurements to include, from multiple available measurements, in estimating structural damage. This is important, since not all the measurements may provide useful information and could reduce performance when processed. Within the adaptive learning framework, the data selection problem is formulated to choose those measurements which are most representative of the diversity within a damage state. This is achieved by extracting time-frequency analysis based statistical similarity features from the measurements, and selecting uniformly distributed subsets to build representative reference sets. The utility of the proposed method is demonstrated by improvements in adaptive learning performance for the estimation of fatigue damage in an aluminum compact tension sample.