(Signalklassificering). Among other things, a comparison was conducted of the state-of-the-art neural network-based modulation classifiers available at the time. Since large quantities of reliable data are extremely important for deep learning, and published datasets for modulation classification have limitations, we have generated our own datasets. The datasets can be designed to include a variety of imperfections that the classifiers should be robust against, such as channel fading, frequency offset, and suboptimal sampling. We show that the networks are robust to most imperfections, but that channel fading, apart from white noise, is the imperfection that is most difficult for the networks to handle. We also conduct a smaller study investigating how the networks can be adapted to accommodate more classes than those present during the initial training. The results show that fine-tuning is required in order to classify signals from the new classes