A machine learning dataset for FRB detection in raw data

Abstract: We introduce a machine learning FRB dataset that can train the ML algorithms to reach the FRBs in raw data. It has 8020 FRB simulation images, 4010 non-FRB and 4010 RFI simulation images built from the public FRB observations, and can be expanded in any number as needed. This work provides an open-source dataset for state of art AI to the comparison of FRB event recognition algorithms. The dataset provides image and NumPy format files for both convolutional neural networks and classic machine learning algorithms. The dataset can implement FRB/non-FRB classification, or FRB/RFI/Blank classification. In the example, we used 31 pre-trained classic CNNs. In FRB/non-FRB classification, it achieves the accuracy of 90-92% in the first training epoch and max accuracy of 99.8% in real FRB dataset testing.