PUB - Publikationen an der Universität Bielefeld

Pulsars are rotating neutron stars which emit faint beams of electromagnetic radiation. Astronomers are able to observe these beams as faint, dispersed pulses. In pulsar searches large effort is expended to discover these pulses in time- and frequency-resolved data from radio telescopes. Simultaneously recovering the frequency-dependent delay (dispersion) and the periodicity of the signal is a complex and demanding task, which is further exacerbated by the presence of various types of radio-frequency interference (RFI) and observing-system effects. New observing systems provide higher bandwidths, higher data volumes and greater overall sensitivity (also to RFI), which further enhances these challenges.

A novel approach for the analysis of pulsar search data is presented in this thesis. I developed and neural-network-based pipeline capable of correcting for the (*a priori* unknown) interstellar dispersion while suppressing a wide range of RFI signals and system effects. A convolutional neural network using dilated convolutions dedisperses pulsar pulses. The classification is performed by classifiers combining standard algorithms for periodicity searches such as the Fast Fourier Transform (FFT) and the Fast Folding Algorithm (FFA) with convolutional layers. This architecture can be trained in an end-to-end manner to identify faint pulses with an unknown amount of dispersion. The performance of the model relies heavily on the training process. Optimising this training process and allowing the model to have the best possible performance is a big focus of this work.

I developed an approach to train the network with successively weaker signals which enables the network to detect strong pulsars after a small amount of time and increase the sensitivity in subsequent training steps. The training process is guided by multiple loss functions which help the network to build useful representations of the input data. The training of the model is performed using simulated pulsars and real pulsar survey observations. I use a technique to relabel the used survey observations which allows training the model on a pulsar survey without labelling the pulsars in the survey first. I am able to detect a wide range of real pulsars and the dedispersion of my neural network is competitive with knowing the pulsar dispersion already and performing conventional dedispersion.