Extreme Mass Ratio Inspirals (EMRI) are one of the most interesting signals which will be detected by Laser Interferometer Space Antenna (LISA). They will sample highly curved spacetime around Massive Black Holes (MBH) and allow for insight into the strong regime of gravity. EMRIs are signals which come from the compact objects falling onto MBH, they typically last for 10^5 number of cycles which translates into an order of one year length. They have relatively low signal-to-noise ratios (SNR) and typically it is required to observe them for a long time to be able to accumulate enough SNR. The signal has a very complicated structure as it contains the evolution of the harmonics of three fundamental frequencies: orbital frequency, perihelion precession frequency and orbital plane precession frequency. At the moment the best up-todate approaches for the EMRI parameters estimation involve semi-coherent searches, which are based on the combination of analyses for relatively short segments of data. To facilitate attempts of applying other machine learning driven approaches to parameter estimation, such as Normalising Flows, we have to find the best way to represent a signal in the compressed way. Therefore this project will be focused on investigating different ways to accumulate information from long lasting signals. We will look at different types of Autoencoders, starting from linear Principle Component decomposition and simple fully connected networks, gradually increasing it to more complicated methods. The goal in the end of the project is to find the most compact representation of the data which will preserve sufficient amount of information.