Aging bridge infrastructure appears to become a major challenge in many industrialized countries. Numerous bridges are in bad condition and the current pace of repair and replacement as well as the available financial resources hence demand for a reliable bridge monitoring to facilitate an extended operation period of existing bridges. Nowadays, prestressed concrete bridges are prevalent among other construction types but may suffer from stress corrosion cracking of steel tendons. To detect wire breaks in bridge tendons, recent research suggests the use of acoustic emission analysis. In this work, we propose the use of semi-supervised learning techniques for anomaly detection to detect wire breaks in tendons of prestressed concrete bridges. Particularly, we utilize only acoustic emissions due to traffic and other environmental influences, recorded on a real bridge in operation, to initialize the local outlier factor algorithm. We then apply the initialized local outlier factor algorithm to two separate datasets with more than 500 wire break signals recorded on two different types of bridge girders. It is shown that the anomaly-based approach outperforms a supervised k-nearest neighbors classifier trained using wire breaks from only one girder. An evaluation on the wire break signals from the second bridge girder, not seen during the training phase, shows an improvement of the average recall score from 38 % to more than 99 % for the anomaly-based approach compared to the supervised k-nearest neighbors classifier. Considering the diversity of bridge constructions and the fact that availability of acoustic emission signals due to wire breaks is limited, semi-supervised learning seems to be a suitable approach for wire break detection. Furthermore, acoustic emissions due to normal environmental and operational conditions could be easily and cost-effectively recorded during an initialization phase of any monitoring system and thus be utilized to initialize an anomaly detector for each specific infrastructure.