The impedance-based structural health monitoring (SHM) method has come to the forefront in the SHM community due to its practical potential for real applications. In the impedance-based SHM method, the selection of optimal frequency ranges plays an important role in improving the sensitivity of damage detection, since an improper frequency range can lead to erroneous damage detection results and provide false positive damage alarms. To tackle this issue, this paper proposes an innovative technique for autonomous selection of damage-sensitive frequency ranges using artificial neural networks (ANNs). First, the impedance signals are obtained in a wide frequency band, and the signals are split into multiple sub-ranges of this wide band. Then, the predefined damage index is evaluated for each sub-range by comparing impedance signals between the intact and the concurrent cases. Here, the cross correlation coefficients (CCs) are used as the predefined damage index. The ANN is constructed and trained using all CC values at multiple frequency ranges as multi-inputs and the real damage severity as the single output for various preselected damage scenarios, so that subsequent damage estimations may be carried out by selecting the governing frequency ranges autonomously. The performance of the proposed approach has been examined via a series of experimental studies to detect loose bolts and cracks induced on real steel bridge and building structures. It is found that the proposed approach autonomously determines the damage-sensitive frequency ranges and can be used for effective evaluation of damage severity in a wide variety of damage cases in real structures.