Predictive maintenance is an anticipatory maintenance strategy designed to predict impending errors or the remaining useful lifetime of a technical system. With its predictions, maintenance actions can be taken condition-based and proactively leading to an efficient use of resources. The basis for this approach is the recording of extensive machine data and their analysis with means of machine learning to learn complex relationships between condition data and related target variables, e. g. the remaining useful life. However, the existence of numerous monitoring data does not imply that there is a good information base for the development of meaningful predictive models. In this paper, for example, a study is presented where the challenge was to predict the optimal time to replace a wear-induced tool of a milling machine. Since in the past the milling tools were often replaced way before the end of their actual lifetime, a predictive model based on this training data would describe a risk averse and thus non-optimal maintenance strategy. Against this background, the present paper deals with the development of a methodology for optimizing a maintenance strategy based on incomplete information from industrial practice. The proposed methodology is based on a broad range of analytical techniques, including feature extraction for the preprocessing of time series data, unsupervised learning for clustering condition monitoring data and a recurrent neural network for the development of a predictive model to determine the remaining useful life. With the approach developed, it is possible to replace decisions which were so far taken based on subjective criteria with data-driven decisions to increase the milling tool lifetime. Additionally, this work contributes to the question how to develop decision support systems despite vague and insufficient information with the means of machine learning.