PSRFINET: Radio Frequency Interference Detection in Pulsar Data with Deep Residual Networks

Radio Frequency Interference (RFI) is a hindrance to high-precision pulsar timing experiments aimed at detecting the stochastic gravitational wave background. Thresholds set by linear combinations of statistical quantities are among the most common approaches to RFI flagging of folded pulse profiles. We propose a deep convolutional neural network approach to RFI flagging called PSRFINET that treats two-dimensional arrays of pulse profiles (rotational phase versus radio frequency) as images and performs feature learning on labelled RFI samples. We train and validate multiple deep residual neural networks on many hours of pulsar observations (thousands of 8 second sub-integrations) of MeerKAT L-band data where the ground truth is generated from Clfd and Coastguard software packages for RFI mitigation. A method of combining the separate ground truths aimed at enhancing the RFI mitigation capabilities of the networks is also explored. The performance of the networks was evaluated by examining the classification metrics of area under the curve of the receiver operating characteristic (AUROC), Precision-Recall (PR) and F1 scores. Our preliminary results show an AUROC of more than 0.91 and PR of 0.67 which indicates that although the neural networks are capable of distinguishing between clean and corrupted frequency channels, precision and recall scores are limited by a class imbalance of a small amount of RFI with respect to clean channels. We also discuss our approach to develop a statistical objective figure of merit for evaluating and comparing the effectiveness of different RFI flagging approaches in the data.