Indoor localization has attracted increasing attention due to demands for location awareness, where GPS (global positioning system) does not work. Fingerprint-based indoor localization is one of standard approaches, but it does not gain recognition as a practical technique due to labor-intensive database maintenance. To protect the degradation of localization accuracy caused by the environmental variations such as relocation of APs (access points) and broken-down APs, this paper proposes a new unlabeled maintenance approach to detect variations of fingerprint pattern, where a massive number of unlabeled data can be easily collected in contrast to costly labeled data. The proposed variation detection algorithm, in offline phase, is based on feature extraction and machine learning techniques. In online phase, KL divergence by information theory is used to measure similarity between existing database and up-to-date data. For experimental study, Wi-Fi RSSI (received signal strength indicator) data including labeled data and unlabeled data are collected from the multi-story office building. The proposed algorithm validates (1) clear relationship between KL divergence metric and positioning accuracy, (2) ability of distinguishing minor and major variations, and (3) accurate variation detection based on unlabeled data.