The installation date is often a primary consideration to explain the deterioration of the sanitation system. With this knowledge, the managers can (through deterioration models) predict the current condition of pipes that have not yet been examined, which is an essential information to take decision in a limited budget context. The data that has to be dealt with has different levels of complexity. The data has heterogeneous sources, a large volume, and limited information on its labelling (years): only 24% of the linear amount is known for the sanitation network. The underlying database will consist of the known features of the pipes (geometric profile, installation depth and so on). This study assessed the effect of some semi-supervised learning methods, and proposed a new approach suitable for this type of data.