This thesis investigates the optimization of autoencoders for the classification of multivariate real-time measurement data emanating from production machinery. The objective is to detect anomalous time-series samples that may indicate process failures. The classification performance depends on a variety of factors, the most important of which were analyzed in separate series of experiments. The autoencoders were set up using an existing framework that allows the inclusion of uni- or bidirectional Long Short-Term Memory (LSTM) layers in order to track long-term dependencies in the data. The error between the original signal and the corresponding reconstruction obtained by the autoencoder was used as a measure of the degree a sample is believed to be anomalous. Via an error threshold, the data samples were classified as either anomalous or non-anomalous. Since the distribution over the reconstruction errors of different samples was right-skewed, a skewness-adjusted threshold setting was performed. In a series of tests, autoencoders with different architectures were compared with regard to their suitability for anomaly detection. Further experiments involved the use of several techniques for initializing the weight parameters. In addition, various methods for optimizing the most impactful hyperparameters were evaluated. Considering the results of the experiments mentioned above, separate autoencoders for the two main phases of the monitored process were optimized and tested. These models were combined with a statistical outlier detection tool based on key performance indicators in order to form hybrid learning models. The data set analyzed in the experiments was gathered from instrumented machinery used in a ground improvement process for building foundations. The aim of this thesis was to support ongoing research at the Chair of Automation involving the use of machine learning techniques in combination with classical methods for machine data analysis.