Monitoring industrial processes are typical tasks for human maintenance experts. Unfortunately, this kind of expert needs a high amount of domain knowledge and is thus very rare. This leads to circumstances in which a frequent monitoring of high dimensional sensor data is desired but cannot be implemented in the long run. In the past, multiple approaches based on signal similarity or prediction models, have been proposed. Within this contribution we try to transfer knowledge from Recurrent Neural Network (RNN)-based speech translation techniques onto bearing fault diagnosis. Therefore, we use a Long-Short-Term Memory (LSTM)-Autoencoder-based system to extract features from raw time series data and receive information about the systems current health state. We also evaluate the learned representations for different bearing damages and propose an extention to our model based on an Attention-LSTM approach which let the network decide which parts of the sequence are relevant to look at.