logo

SCIENCE CHINA Physics, Mechanics & Astronomy, Volume 62, Issue 5: 959507(2019) https://doi.org/10.1007/s11433-018-9388-3

Pulsar candidate selection using ensemble networks for FAST drift-scan survey

More info
  • ReceivedDec 7, 2018
  • AcceptedMar 8, 2019
  • PublishedMar 20, 2019
PACS numbers

Abstract

The Commensal Radio Astronomy Five-hundred-meter Aperture Spherical radio Telescope (FAST) Survey (CRAFTS) utilizes the novel drift-scan commensal survey mode of FAST and can generate billions of pulsar candidate signals. The human experts are not likely to thoroughly examine these signals, and various machine sorting methods are used to aid the classification of the FAST candidates. In this study, we propose a new ensemble classification system for pulsar candidates. This system denotes the further development of the pulsar image-based classification system (PICS), which was used in the Arecibo Telescope pulsar survey, and has been retrained and customized for the FAST drift-scan survey. In this study, we designed a residual network model comprising 15 layers to replace the convolutional neural networks (CNNs) in PICS. The results of this study demonstrate that the new model can sort $>$96% of real pulsars to belong the top 1% of all candidates and classify $>$1.6 million candidates per day using a dual-GPU and 24-core computer. This increased speed and efficiency can help to facilitate real-time or quasi-real-time processing of the pulsar-search data stream obtained from CRAFTS. In addition, we have published the labeled FAST data used in this study online, which can aid in the development of new deep learning techniques for performing pulsar searches.


Acknowledgment

This work was supported by the National Key Research and Development Program of China (Grant No. 2017YFA0402600), the Natural Science Foundation of Shandong (Grant No. ZR2015FL006), the CAS International Partnership Program (Grant No. 114A11KYSB20160008), the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB23000000), the National Natural Science Foundation of China (Grant Nos. 61472043, 11743002, 11873067, 11690024, and 11725313), the Joint Research Fund in Astronomy (Grant No. U1531242) under Cooperative Agreement between the NSFC and CAS and National Natural Science Foundation of China (Grant No. 11673005), and the Chinese Academy of Sciences Pioneer Hundred Talents Program. The authors also thank Chavonne Bowen and Alan Ho for labeling FAST pulsar candidates, the PALFA, GBNCC team and the Arecibo Remote Command Center (ARCC) students, Cherry Ng, Meng Yu, et al. for labeling and sharing their data. The authors thank the referee's constructive comments and suggestions.


References

[1] Manchester R. N., Lyne A. G., Camilo F., Bell J. F., Kaspi V. M., D'Amico N., McKay N. P. F., Crawford F., Stairs I. H., Possenti A., Kramer M., Sheppard D. C.. Mon. Not. R. Astron. Soc., 2001, 328: 17 CrossRef ADS Google Scholar

[2] Burke-Spolaor S., Bailes M., Johnston S., Bates S. D., Bhat N. D. R., Burgay M., D'Amico N., Jameson A., Keith M. J., Kramer M., Levin L., Milia S., Possenti A., Stappers B., van Straten W.. Mon. Not. R. Astron. Soc., 2011, 416: 2465 CrossRef ADS arXiv Google Scholar

[3] Deneva J. S., Cordes J. M., McLaughlin M. A., Nice D. J., Lorimer D. R., Crawford F., Bhat N. D. R., Camilo F., Champion D. J., Freire P. C. C., Edel S., Kondratiev V. I., Hessels J. W. T., Jenet F. A., Kasian L., Kaspi V. M., Kramer M., Lazarus P., Ransom S. M., Stairs I. H., Stappers B. W., van Leeuwen J., Brazier A., Venkataraman A., Zollweg J. A., Bogdanov S.. Astrophys. J., 2009, 703: 2259 CrossRef ADS arXiv Google Scholar

[4] Boyles J., Lynch R. S., Ransom S. M., Stairs I. H., Lorimer D. R., McLaughlin M. A., Hessels J. W. T., Kaspi V. M., Kondratiev V. I., Archibald A., Berndsen A., Cardoso R. F., Cherry A., Epstein C. R., Karako-Argaman C., McPhee C. A., Pennucci T., Roberts M. S. E., Stovall K., van Leeuwen J.. Astrophys. J., 2013, 763: 80 CrossRef ADS arXiv Google Scholar

[5] Stovall K., Lynch R. S., Ransom S. M., Archibald A. M., Banaszak S., Biwer C. M., Boyles J., Dartez L. P., Day D., Ford A. J., Flanigan J., Garcia A., Hessels J. W. T., Hinojosa J., Jenet F. A., Kaplan D. L., Karako-Argaman C., Kaspi V. M., Kondratiev V. I., Leake S., Lorimer D. R., Lunsford G., Martinez J. G., Mata A., McLaughlin M. A., Roberts M. S. E., Rohr M. D., Siemens X., Stairs I. H., van Leeuwen J., Walker A. N., Wells B. L.. Astrophys. J., 2014, 791: 67 CrossRef ADS arXiv Google Scholar

[6] Coenen T., van Leeuwen J., Hessels J. W. T., Stappers B. W., Kondratiev V. I., Alexov A., Breton R. P., Bilous A., Cooper S., Falcke H., Fallows R. A., Gajjar V., Grie?meier J. M., Hassall T. E., Karastergiou A., Keane E. F., Kramer M., Kuniyoshi M., Noutsos A., Os?owski S., Pilia M., Serylak M., Schrijvers C., Sobey C., ter Veen S., Verbiest J., Weltevrede P., Wijnholds S., Zagkouris K., van Amesfoort A. S., Anderson J., Asgekar A., Avruch I. M., Bell M. E., Bentum M. J., Bernardi G., Best P., Bonafede A., Breitling F., Broderick J., Brüggen M., Butcher H. R., Ciardi B., Corstanje A., Deller A., Duscha S., Eisl?ffel J., Fender R., Ferrari C., Frieswijk W., Garrett M. A., de Gasperin F., de Geus E., Gunst A. W., Hamaker J. P., Heald G., Hoeft M., van der Horst A., Juette E., Kuper G., Law C., Mann G., McFadden R., McKay-Bukowski D., McKean J. P., Munk H., Orru E., Paas H., Pandey-Pommier M., Polatidis A. G., Reich W., Renting A., R?ttgering H., Rowlinson A., Scaife A. M. M., Schwarz D., Sluman J., Smirnov O., Swinbank J., Tagger M., Tang Y., Tasse C., Thoudam S., Toribio C., Vermeulen R., Vocks C., van Weeren R. J., Wucknitz O., Zarka P., Zensus A.. Astron. Astrophys., 2014, 570: A60 CrossRef ADS arXiv Google Scholar

[7] Lynch R. S., Boyles J., Ransom S. M., Stairs I. H., Lorimer D. R., McLaughlin M. A., Hessels J. W. T., Kaspi V. M., Kondratiev V. I., Archibald A. M., Berndsen A., Cardoso R. F., Cherry A., Epstein C. R., Karako-Argaman C., McPhee C. A., Pennucci T., Roberts M. S. E., Stovall K., van Leeuwen J.. Astrophys. J., 2013, 763: 81 CrossRef ADS arXiv Google Scholar

[8] Jiang, P., Yue, Y., Gan, H. et al. Commissioning Progress o FAST, Sci. China-Phys. Mech. Astron. 2019(5) in press. Google Scholar

[9] Li D., Wang P., Qian L., Krco M., Dunning A., Jiang P., Yue Y., Jin C., Zhu Y., Pan Z., Nan R.. IEEE Microwave, 2018, 19: 112-119 CrossRef ADS arXiv Google Scholar

[10] Lyon R. J., Stappers B. W., Cooper S., Brooke J. M., Knowles J. D.. Mon. Not. R. Astron. Soc., 2016, 459: 1104 CrossRef ADS arXiv Google Scholar

[11] Keith M. J., Jameson A., Van Straten W., Bailes M., Johnston S., Kramer M., Possenti A., Bates S. D., Bhat N. D. R., Burgay M., Burke-Spolaor S., D'Amico N., Levin L., McMahon P. L., Milia S., Stappers B. W.. Mon. Not. R. Astron. Soc., 2010, 409: 619 CrossRef ADS arXiv Google Scholar

[12] Wang K., Guo P., Yu F., Duan L., Wang Y., Du H.. Inter. J. Comput. Intel. Syst., 2018, 11: 575 CrossRef Google Scholar

[13] Eatough R. P., Molkenthin N., Kramer M., Noutsos A., Keith M. J., Stappers B. W., Lyne A. G.. Mon. Not. R. Astron. Soc., 2010, 407: 2443 CrossRef ADS arXiv Google Scholar

[14] Lee K. J., Stovall K., Jenet F. A., Martinez J., Dartez L. P., Mata A., Lunsford G., Cohen S., Biwer C. M., Rohr M., Flanigan J., Walker A., Banaszak S., Allen B., Barr E. D., Bhat N. D. R., Bogdanov S., Brazier A., Camilo F., Champion D. J., Chatterjee S., Cordes J., Crawford F., Deneva J., Desvignes G., Ferdman R. D., Freire P., Hessels J. W. T., Karuppusamy R., Kaspi V. M., Knispel B., Kramer M., Lazarus P., Lynch R., Lyne A., McLaughlin M., Ransom S., Scholz P., Siemens X., Spitler L., Stairs I., Tan M., van Leeuwen J., Zhu W. W.. Mon. Not. R. Astron. Soc., 2013, 433: 688 CrossRef ADS arXiv Google Scholar

[15] Bates S. D., Bailes M., Barsdell B. R., Bhat N. D. R., Burgay M., Burke-Spolaor S., Champion D. J., Coster P., D'Amico N., Jameson A., Johnston S., Keith M. J., Kramer M., Levin L., Lyne A., Milia S., Ng C., Nietner C., Possenti A., Stappers B., Thornton D., van Straten W.. Mon. Not. R. Astron. Soc., 2012, 427: 1052 CrossRef ADS arXiv Google Scholar

[16] Morello V., Barr E. D., Bailes M., Flynn C. M., Keane E. F., van Straten W.. Mon. Not. R. Astron. Soc., 2014, 443: 1651 CrossRef ADS arXiv Google Scholar

[17] P. Guo, F. Duan, P. Wang, Y. Yao, and X. Xin,. arXiv Google Scholar

[18] L. Connor, and J. van Leeuwen,. arXiv Google Scholar

[19] Zhang Y. G., Gajjar V., Foster G., Siemion A., Cordes J., Law C., Wang Y.. Astrophys. J., 2018, 866: 149 CrossRef ADS arXiv Google Scholar

[20] Gajjar V., Siemion A. P. V., Price D. C., Law C. J., Michilli D., Hessels J. W. T., Chatterjee S., Archibald A. M., Bower G. C., Brinkman C., Burke-Spolaor S., Cordes J. M., Croft S., Enriquez J. E., Foster G., Gizani N., Hellbourg G., Isaacson H., Kaspi V. M., Lazio T. J. W., Lebofsky M., Lynch R. S., MacMahon D., McLaughlin M. A., Ransom S. M., Scholz P., Seymour A., Spitler L. G., Tendulkar S. P., Werthimer D., Zhang Y. G.. Astrophys. J., 2018, 863: 2 CrossRef ADS arXiv Google Scholar

[21] Zhu W. W., Berndsen A., Madsen E. C., Tan M., Stairs I. H., Brazier A., Lazarus P., Lynch R., Scholz P., Stovall K., Ransom S. M., Banaszak S., Biwer C. M., Cohen S., Dartez L. P., Flanigan J., Lunsford G., Martinez J. G., Mata A., Rohr M., Walker A., Allen B., Bhat N. D. R., Bogdanov S., Camilo F., Chatterjee S., Cordes J. M., Crawford F., Deneva J. S., Desvignes G., Ferdman R. D., Freire P. C. C., Hessels J. W. T., Jenet F. A., Kaplan D. L., Kaspi V. M., Knispel B., Lee K. J., van Leeuwen J., Lyne A. G., McLaughlin M. A., Siemens X., Spitler L. G., Venkataraman A.. Astrophys. J., 2014, 781: 117 CrossRef ADS arXiv Google Scholar

[22] K. He, X. Zhang, S. Ren, and J. Sun,. arXiv Google Scholar

[23] M. Hardt, and T. Ma,. arXiv Google Scholar

[24] S. S. Du, W. Hu, and J. D. Lee,. arXiv Google Scholar

  • Figure 1

    Real pulsar candidate obtained from the FAST drift-scan survey. The time domain displays an intermittent signal caused by the pulsar drifting in and out of the beam. The phase versus time plot reveals a significant subpulse drift in the pulsar.

  • Figure 2

    The example of a non-pulsar candidate obtained from the FAST data. The frequency versus phase plot contains features that resemble the pulsar signals. However, upon close examination, these features become entirely vertical when the plot is refolded with DM set to zero, demonstrating that the signal is not dispersed. Thus, this candidate is most likely to have originated from the ground interference.

  • Figure 3

    Illustration of a residual block. Residual learning: a building block.

  • Figure 4

    (Color online) Diagram of a 15 layered ResNet model. Conv refers to convolutional operation. The “2 Residual Blocks" component contains two building blocks that possess an identical architecture. The input image is $64~\times~64~\times~1$, and the output sizes are $H~\times~W~\times~N$, where $H$ and $W$ denote the height and width of the tensor, and $N$ denotes the number of features.

  • Figure 5

    Diagram of the PICS-ResNet model. The first layer classifies the individual features (the pulse profile, time versus phase plot, frequency versus phase plot, and DM curve), whereas the second layer classifies the candidates based on the results of the first layer. The SVM components represent the support vector machine model, while the ResNet components represent the residual network model. See Figure 4for a schematic of the ResNet model.

  • Figure 6

    (Color online) Convergence of the 15 layers of ResNet over the training process. The $x$-axis represents the training epochs, whereas the $y$-axis represents the F$_1$ score.

  • Figure 7

    (Color online) The PICS-ResNet and PICS learning curves. The $x$-axis represents the training sample set from 2000 to 12000, whereas the $y$-axis represents F$_1$ score.

  • Figure 8

    (Color online) PICS-ResNet and PICS recall curves. The $x$-axis represents the fraction of the candidates examined, whereas the $y$-axis represents the recall rate of the examined candidates. Here, the recall rate is calculated based on the fraction of pulsar signals (including 56 fundamental signals and 221 harmonic signals) ranked in the top $X$ fraction of the candidates.

  • Table 1   Number of positive and negative examples in the datasets
    Dataset names Positive Negative Total
    PALFA 3951 6672 10623
    HTRU 903 271 1174
    FAST 837 998 1835
    GBNCC 277 89731 90008
  • Table 2   Binary classification confusion matrix, which defines all the outcomes of predictions, including true negative (TN), false negative (FN), false positive (FP), and true positive (TP)
    Outcomes Negative prediction Positive prediction
    RFI true negative false positive
    True pulsar false negative true positive
  • Table 3   The number of training samples was continuously increased from 2000 to 12000. Five validation tests were performed for each data point, and the means of the $_1$ score were obtained
    MethodTraining dataset F$_1$ score
    $N$=2000$N$=4000$N$=6000$N$=8000$N$=10000$N$=12000
    PICS training 0.980.980.990.990.990.99
    PICS-ResNet training 0.980.980.980.980.980.98
    PICS validation 0.860.900.910.910.920.92
    PICS-ResNet validation 0.890.910.920.920.910.92
  • Table 4   Classification results of the PICS and PICS-ResNet models using the GBNCC dataset when the models were trained with the PALFA and HTRU data but not the FAST data. Only the candidates who were ranked in the top $1%$ of the dataset were selected, and the recall rates of the pulsars and their harmonics were calculated
    Method Fundamental Harmonic PSR
    (recall)(recall) (recall)
    PICS (top $1%$) 52 ($93%$) 194 ($88%$) 246 ($89%$)
    PICS-ResNet (top $1%$) 51 ($91%$) 190 ($86%$) 241 ($87%$)
  • Table 5   Classification results of the PICS and PICS-ResNet models using the GBNCC dataset. The models were trained with datasets from the PALFA, HTRU, and FAST. Only the candidates ranked in the top $1%$ of the dataset were selected, and the recall rates of the pulsars and their harmonics were calculated
    Method Fundamental Harmonic PSR
    (recall)(recall)(recall)
    PICS (top $1%$) 52 ($93%$) 201 ($91%$) 253 ($91%$)
    PICS-ResNet (top $1%$) 54 ($96%$) 211 ($96%$) 265 ($96%$)
  • Table 6   Number of pulsars identified by each model when tested using the FAST data. The test set comprised a total of 326 pulsar candidates
    Method Recognition pulsar Missing pulsar Recall (%)
    PICS 310 16 $95$
    PICS-ResNet 320 6 $98$

Copyright 2020 Science China Press Co., Ltd. 《中国科学》杂志社有限责任公司 版权所有

京ICP备18024590号-1       京公网安备11010102003388号