Approaches to machine and equipment maintenance based on data analytics and artificial intelligence are trending in modern manufacturing. These methods are used to predict the remaining useful life (RUL) of equipment and thus enable forward maintenance planning. However, for predictive maintenance systems, it is also necessary to detect anomalies in operation and classify the occurring errors. Classical approaches of supervised machine learning are often in this case unusable because those methods require a large amount of run-to-failure data (R2F), which is often not possible to collect due to the undesirable character of failure states in the manufacturing process. The paper presents and tests several methods of detecting device fault states using an autoencoder network, which offers a beneficial solution in the case of the unavailability of R2F data in the system.