Answering with some theoretical understanding of Variational auto-encoders.
In the general architecture of encoders and decoders, the encoder encodes the input a latent-space, and the decoder reconstructs the input from the encoded latent space.
However, the Variational auto-encoders (VAE), the input is encoded to a latent-distribution instead of a point in ...
Splitting a time-series into analysis windows, usually with overlap, is quite common practice. In anomaly detection, but also in classification or forecasting. It works great as long as your anomalies can be detected by analyzing such windows independently.
In such a setup, the length of the analysis window becomes a critical hyper-parameter - and will be ...
I'd expect the encodings to be quite idiosyncratic. It can be trivially proven that a decoder can be trained to produce any single output given a code of , and so I'm not certain that the code your AE will converge to will correspond with the inputs in an especially meaningful way.
From my experience, the cosine similarity loss (tf.keras.losses.CosineSimilarity) works best for text autoencoders with Keras. You can find the documentation at: