Bayesian neural network models have gained an increased popularity recently. They offer a path for constructing generative probabilistic models in settings where classical methods are lacking the ability to process large amounts of high dimensional data. We explore different types of autoencoder based models with respect to the way they enforce a structure in the latent space. These models differ by their latent prior distributions and their loss functions. We apply the models to spectrographic data from a tokamak in which we try to identify the time of occurence and a frequency of a plasma instability. The generative models are used to encode the data from the image domain to a low dimensional latent space in which an additional model is used to identify the anomalous patterns. We present a comparison of different approaches and propose a direction of further work.