This thesis examines the combination of deep learning and statistical data analysis for the unsupervised detection of outliers in unlabelled real-time time-series data originating from sensors and actuators. It is investigated if and when the additional application of machine learning is preferable to classical methods, where its strengths and shortcomings lie. A hybrid approach is introduced to enable an exhaustive and precise identification of anomalies. Its components are based on previous work performed by the Chair of Automation at the University of Leoben, which includes the segmentation of different phases in the industrial process and the definition of key point indicators (KPIs) by accessing physical knowledge and experience. The hybrid model contains the calculation of the interquartile range to define outliers, as well as the training and testing of a hyperparameter-optimized deep learning algorithm. The approach was applied to datasets from machinery used in the construction of anchoring piles for the foundations of buildings.