Addressing Distribution Discrepancies in Pulsar Candidate Identification via Bayesian-neural-network-based Multimodal Incremental Learning

With the advancement of astronomical observation technology and the substantial increase in data volume, traditional methods for pulsar identification are increasingly challenged by the dynamic nature of data distributions. To address this, our study introduces a multimodal incremental learning approach utilizing Bayesian neural networks. This method enables the model to adapt to new data distributions while preserving the knowledge of previous data. In our experiments, we utilized pulsar data sets from two telescopes and compared our new method with traditional techniques. The research results demonstrate that our method performs comparably to traditional methods across all evaluation metrics, while showing a significant improvement in handling data distribution discrepancy, with the F1-score increasing from approximately 70% to over 95%. Specifically, our model achieved an accuracy of 97.93%, a recall of 96.13%, and an F1-score of 97.02% under conditions of distributional disparities. These findings not only confirm the model's capability to adapt to dynamic data environments but also effectively address the challenges of data distribution discrepancy, significantly enhancing the predictive accuracy of pulsar identification in the context of evolving and variable radio frequency interference environments.