Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Identifying Quasi-Periodic Micropulses in Pulsars with FAST Using Convolutional Neural Networks

Shidong Wang, Hui Liu, Ru-Shuang Zhao, Baoqiang Lao, Yong-Kun Zhang, Y. F. Xiao, Pei Wang, Di Li, R. W. Tian, Z. F. Tu, Q. Zhou, Z. J. Zhang, Qijun Zhi, Shijun Dang, Kun Yang

TL;DR

This work tackles the challenge of identifying quasi-periodic micropulses (QMP) in massive FAST pulsar datasets by introducing a Dual-Stage Residual Network (DSR) that fuses information from single-pulse profiles and their ADP spectra. Model 1 (ResNet-18) processes 2D images of individual pulses, while Model 2 (ResNet-34) validates candidates using corresponding ADP images, achieving a combined precision of 0.9585 and recall of 0.9610 on the 2020 PSR B1933+16 test set, with strong cross-year and cross-pulsar performance. The approach outperforms single-stage and 1D baselines, demonstrates robust generalization across multiple pulsars and emission components, and yields detailed $P_\mu$ statistics that are broadly consistent with previous studies while revealing year-to-year bimodalities likely tied to scintillation. The results provide a scalable, automated pipeline for large-scale QMP identification, enabling deeper physical insights into pulsar magnetospheric geometry and emission processes, and highlight avenues for expanding training data and adopting end-to-end architectures for future FAST-era surveys.

Abstract

Quasi-periodic MicroPulses (QMP) are quasi-periodic microstructural features manifested in individual pulsar radio pulses, the study of which is crucial for understanding pulsar radiation mechanisms. Manual identification of QMP in large-scale pulsar single-pulse datasets remains highly inefficient. To address this, we propose a Dual-Stage Residual Network (DSR) that achieves automated QMP detection in FAST observational data through joint analysis of single-pulse profiles and their Amplitude Distribution Profiles (ADP), defined as the power spectra of the autocorrelation function derivatives of the microstructure residuals. The model was trained on PSR B1933+16 data from 2019 (10,486 single pulses) and evaluated on manually annotated PSR B1933+16 data from 2020 (9,657 single pulses). DSR achieved 96.10\% recall and 95.85\% precision on the test set. This approach provides an automated pipeline for large-scale, reproducible QMP identification and establishes the foundation for in-depth investigation of their physical mechanisms.

Identifying Quasi-Periodic Micropulses in Pulsars with FAST Using Convolutional Neural Networks

TL;DR

This work tackles the challenge of identifying quasi-periodic micropulses (QMP) in massive FAST pulsar datasets by introducing a Dual-Stage Residual Network (DSR) that fuses information from single-pulse profiles and their ADP spectra. Model 1 (ResNet-18) processes 2D images of individual pulses, while Model 2 (ResNet-34) validates candidates using corresponding ADP images, achieving a combined precision of 0.9585 and recall of 0.9610 on the 2020 PSR B1933+16 test set, with strong cross-year and cross-pulsar performance. The approach outperforms single-stage and 1D baselines, demonstrates robust generalization across multiple pulsars and emission components, and yields detailed statistics that are broadly consistent with previous studies while revealing year-to-year bimodalities likely tied to scintillation. The results provide a scalable, automated pipeline for large-scale QMP identification, enabling deeper physical insights into pulsar magnetospheric geometry and emission processes, and highlight avenues for expanding training data and adopting end-to-end architectures for future FAST-era surveys.

Abstract

Quasi-periodic MicroPulses (QMP) are quasi-periodic microstructural features manifested in individual pulsar radio pulses, the study of which is crucial for understanding pulsar radiation mechanisms. Manual identification of QMP in large-scale pulsar single-pulse datasets remains highly inefficient. To address this, we propose a Dual-Stage Residual Network (DSR) that achieves automated QMP detection in FAST observational data through joint analysis of single-pulse profiles and their Amplitude Distribution Profiles (ADP), defined as the power spectra of the autocorrelation function derivatives of the microstructure residuals. The model was trained on PSR B1933+16 data from 2019 (10,486 single pulses) and evaluated on manually annotated PSR B1933+16 data from 2020 (9,657 single pulses). DSR achieved 96.10\% recall and 95.85\% precision on the test set. This approach provides an automated pipeline for large-scale, reproducible QMP identification and establishes the foundation for in-depth investigation of their physical mechanisms.
Paper Structure (16 sections, 2 equations, 10 figures)

This paper contains 16 sections, 2 equations, 10 figures.

Table of Contents

  1. Introduction
  2. Data
  3. Observational Data Processing
  4. QMP Classification Methodology
  5. Image Generation and Preprocessing
  6. Dual-Stage Residual Network Model
  7. Experimental Results and Analysis
  8. Performance Evaluation Metrics
  9. Test Results
  10. Discussion
  11. Comparison of 1D Data and Images as Model Input
  12. Analysis of Performance Discrepancies in Single-Stage Models
  13. Attribution of Missed and False Detections in the DSR Model
  14. Distribution Characteristics of the Quasi-periodic Micropulse Period $P_\mu$
  15. Astrophysical Implications of Bimodal Quasi-periodic Distributions
  16. ...and 1 more sections

Figures (10)

  • Figure 1: Diagnostic plot comparison of QMP and NQMP in PSR B1933+16. (i) Pulse 789 with QMP: (a) Original subpulse data (black trace) exhibits pronounced and clustered quasi-periodic microstructure; (b) ACF of microstructure features (blue trace) displays significant periodicity; (c,d) Power spectra (PSD and ADP) show distinct unimodal signatures. (ii) Pulse 1248 with NQMP: (a) Original subpulse data (black trace) shows random fluctuations with indistinct microstructure; (b) ACF of microstructure features (blue trace) presents chaotic patterns; (c,d) Power spectra lack significant unimodal features. (iii) Pulse 892 with NQMP: (a) Original subpulse data (black trace) exhibits microstructure with clear oscillations; (b) Blue trace shows no periodicity; (c,d) Power spectra lack significant unimodal characteristics. (iv) Pulse 773 with NQMP: (a) Original subpulse data (black trace) lacks pronounced clustered microstructure, but contains intrinsic profile oscillations that were misidentified as QMP during microstructure extraction (blue line); thus (b) blue trace exhibits periodicity and (c,d) power spectra show significant unimodal features.
  • Figure 2: Representative examples of positive and negative samples for single-pulse images. Top row: three positive samples (QMPs); bottom row: three negative samples (NQMPs).
  • Figure 3: Characteristic examples of positive and negative samples for ADP-derived power spectra. Top row: three positive samples (QMPs) exhibiting distinct unimodal signatures; bottom row: three negative samples (NQMPs) showing featureless spectra.
  • Figure 4: Schematic diagram of QMP identification by the DSR system. Model 1 and Model 2 serve as conditional judgment mechanisms that determine classification by comparing whether an image's positive-class prediction probability exceeds its negative-class probability: if yes, it is classified as positive; otherwise as negative. The input consists of single-pulse test images, and the final output comprises ADP images matched with QMPs.
  • Figure 5: Workflow of the DSR model: (a) ResNet-18 processes single-pulse images; (b) ResNet-34 processes ADP images. Common components: initial convolutional layer (yellow), residual block groups, global average pooling (brown bar), fully connected layer with Softmax activation (green sphere), and generation of 2D arrays (blue blocks). For ResNet-18, each residual block group contains 2 residual blocks; ResNet-34 contains 3, 4, 6, and 3 blocks per group respectively.
  • ...and 5 more figures