pith. sign in

arxiv: 1906.10947 · v2 · pith:QD6BSWESnew · submitted 2019-06-26 · 🌌 astro-ph.IM · astro-ph.HE

Advanced Signal Reconstruction in Tunka-Rex with Matched Filtering and Deep Learning

P. Bezyazeekov , N. Budnev , O. Fedorov , O. Gress , O. Grishin , A. Haungs , T. Huege , Y. Kazarina
show 16 more authors
M. Kleifges D. Kostunin E. Korosteleva L. Kuzmichev V. Lenok N. Lubsandorzhiev S. Malakhov T. Marshalkina R. Monkhoev E. Osipova A. Pakhorukov L. Pankov V. Prosin F. G. Schr\"oder D. Shipilov A. Zagorodnikov
This is my paper

Pith reviewed 2026-05-25 15:25 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.HE
keywords Tunka-Rexsignal reconstructionmatched filteringautoencoderdeep learningcosmic raysradio detectionair showers
0
0 comments X

The pith

Matched filtering and autoencoder neural networks reconstruct Tunka-Rex radio pulses at lower signal-to-noise ratios than prior methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that matched filtering and deep neural networks built with an autoencoder architecture improve reconstruction of short radio pulses from air showers in the Tunka-Rex array. Conventional reconstruction loses accuracy when background dominates the recorded traces, which restricts detection of weaker events. The new methods recover pulse position and amplitude more reliably, lowering the effective detection threshold and raising the number of usable cosmic-ray events. This matters for studying ultra-high energy cosmic rays because more events yield better statistics on their energy spectrum and arrival directions. The authors compare the techniques directly and present the first Tunka-Rex reconstruction performed by a neural network.

Core claim

Matched filtering using signal templates and an autoencoder neural network can extract the position and amplitude of short radio pulses from background-dominated traces recorded by Tunka-Rex, thereby lowering the detection threshold and increasing analysis efficiency for cosmic-ray air-shower events.

What carries the argument

Matched filtering with pulse templates combined with an autoencoder neural network that processes time-domain traces to recover signal parameters.

If this is right

  • Detection threshold for Tunka-Rex cosmic-ray events is reduced.
  • Analysis efficiency rises because more traces yield usable reconstructions.
  • Signal parameters extracted at low signal-to-noise ratios become more reliable.
  • A working example of deep neural network reconstruction is demonstrated on Tunka-Rex data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same reconstruction pipeline could be transferred to other sparse radio arrays to raise their collected event statistics.
  • Real-time implementation of these filters might allow earlier rejection of noise-only triggers during data taking.
  • Further tests on traces from different antenna spacings would show how much the gain depends on array geometry.

Load-bearing premise

The matched-filter templates and autoencoder can recover true pulse position and amplitude from noise-dominated traces without adding systematic offsets or distortions.

What would settle it

A test that injects simulated pulses of known amplitude and timing into real Tunka-Rex background traces at decreasing signal strengths and checks whether the reconstructed values remain unbiased relative to the injected inputs.

Figures

Figures reproduced from arXiv: 1906.10947 by A. Haungs, A. Pakhorukov, A. Zagorodnikov, D. Kostunin, D. Shipilov, E. Korosteleva, E. Osipova, F. G. Schr\"oder, L. Kuzmichev, L. Pankov, M. Kleifges, N. Budnev, N. Lubsandorzhiev, O. Fedorov, O. Gress, O. Grishin, P. Bezyazeekov, R. Monkhoev, S. Malakhov, T. Huege, T. Marshalkina, V. Lenok, V. Prosin, Y. Kazarina.

Figure 1
Figure 1. Figure 1: Average of 400 Tunka-Rex traces centered at the pulse. The black line indi￾cates the mean value, the shaded area indicates one standard deviation. The expected reduction factor of the noise by averaging is √ 400 = 20, however it is significantly less (about 5), moreover one can see prominent noise feature after the pulse. lowing formula: SNR = S 2 /N2 , (1) where S is the amplitude of the peak of the signa… view at source ↗
Figure 2
Figure 2. Figure 2: Left: Distribution of MF values √ Acc applied on traces of pure noise without air-shower signal. The detection threshold (red line) is set to have < 5% of false positive detections. Right: Correlation between √ Acc and the amplitude of the simulated signal. As was discussed in Ref. [6], the simulated signals reproduce real ones with satisfactory accuracy. As shown below, the methods developed for simulated… view at source ↗
Figure 3
Figure 3. Figure 3: Performance of different configurations of the autoencoder (see main text for explanation). Left: efficiency (fraction of reconstructed events Nrec/Ntot). The rela￾tively low efficiency is due to the detection threshold of the AE at the very low SNRs. Right: purity (fraction of events with properly reconstructed peak position, namely < 5 ns from true value: Nhit/Nrec). The reconstruction quality of the AE … view at source ↗
Figure 4
Figure 4. Figure 4: Examples of the AE performance on simulated data (from top to bottom): 1) correct identification (true positive, signal is perfectly denoised); 2) no identifica￾tion (false negative, due to heavy distortion by the noise), 3) double identification (true+false positive, passing signal-like RFI) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the performance of different methods on simulated data. Left: distribution of the reconstructed events as function of the energy for the standard, MF and AE pipelines; right: angular resolution using the different methods. encoding layer is defined as follows: Si = Smin × 2 D−i , ni = 2i+N−1 , (4) where Si is the size of the i-th filter, ni is the number of filters per layer; Smin = 16 is the… view at source ↗
Figure 6
Figure 6. Figure 6: Example of the autoencoder performance on a measured Tunka-Rex trace showing successful denoising of the typical RFI after the signal, cf. Fig. (1). us to check the performance of MF and AE integrated in Tunka-Rex reconstruc￾tion pipeline, and to use the identical benchmark for comparison between the standard signal reconstruction, matched filtering, and denoising with the au￾toencoder. As metrics we selec… view at source ↗
read the original abstract

The Tunka Radio Extension (Tunka-Rex) is a digital antenna array operating in the frequency band of 30-80 MHz, measuring the radio emission of air-showers induced by ultra-high energy cosmic rays. Tunka-Rex is co-located with the TAIGA experiment in Siberia and consists of 63 antennas, 57 of them in a densely instrumented area of about 1km2. The signals from the air showers are short pulses, which have a duration of tens of nanoseconds and are recorded in traces of about 5{\mu}s length. The Tunka-Rex analysis of cosmic-ray events is based on the reconstruction of these signals, in particular, their positions in the traces and amplitudes. This reconstruction suffers at low signal-to-noise ratios, i.e. when the recorded traces are dominated by background. To lower the threshold of the detection and increase the efficiency, we apply advanced methods of signal reconstruction, namely matched filtering and deep neural networks with autoencoder architecture. In the present work we show the comparison between the signal reconstructions obtained with these techniques, and give an example of the first reconstruction of the Tunka-Rex signals obtained with a deep neural networks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper describes the Tunka-Rex radio array for cosmic-ray air-shower detection and applies matched filtering together with an autoencoder-based deep neural network to reconstruct pulse position and amplitude in background-dominated traces. It claims these techniques lower the detection threshold and increase analysis efficiency relative to standard methods, and presents a comparison of the reconstructions plus one example of DNN output on real data.

Significance. If the claimed improvement in low-SNR reconstruction holds without introducing uncontrolled bias, the work would enlarge the usable event sample for Tunka-Rex and similar sparse radio arrays, directly benefiting energy and composition studies of ultra-high-energy cosmic rays.

major comments (2)
  1. [Abstract] Abstract: the central claim that matched filtering and the autoencoder lower the detection threshold rests on the premise that both recover pulse parameters without systematic offsets larger than the statistical gain; no bias (mean residual vs. injected SNR), pull distributions, or efficiency curves versus SNR are reported for either method on simulated or real low-SNR events.
  2. [Abstract / methods description] The description of the DNN architecture and training does not specify the loss function, regularization, or validation procedure used to ensure that the network does not shift the reconstructed amplitude or timing at the lowest SNRs where the threshold reduction is claimed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments correctly identify gaps in the quantitative validation of the reconstruction methods and in the description of the DNN training procedure. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that matched filtering and the autoencoder lower the detection threshold rests on the premise that both recover pulse parameters without systematic offsets larger than the statistical gain; no bias (mean residual vs. injected SNR), pull distributions, or efficiency curves versus SNR are reported for either method on simulated or real low-SNR events.

    Authors: We agree that explicit demonstration of negligible bias is required to substantiate the threshold-lowering claim. The current manuscript shows qualitative comparisons and one real-data example but does not contain bias plots, pull distributions, or efficiency curves versus SNR. In the revision we will add these metrics for both methods, computed on simulated traces with injected pulses across a range of SNRs. For real data we will explicitly note the absence of independent ground truth and limit claims to consistency checks with the standard reconstruction. revision: yes

  2. Referee: [Abstract / methods description] The description of the DNN architecture and training does not specify the loss function, regularization, or validation procedure used to ensure that the network does not shift the reconstructed amplitude or timing at the lowest SNRs where the threshold reduction is claimed.

    Authors: The referee is correct that these training details are missing. We will expand the methods section to state the loss function (mean-squared error on pulse amplitude and timing), regularization (dropout layers), and validation protocol (separate simulated training/validation/test sets with monitoring of mean residuals on the validation set at low SNR). These additions will document that no systematic offsets were introduced at the SNRs relevant to the claimed threshold improvement. revision: yes

Circularity Check

0 steps flagged

No circularity: methods applied to external data without self-referential derivation

full rationale

The paper applies matched filtering and autoencoder DNNs to Tunka-Rex traces for pulse reconstruction. No derivation chain, fitted-input predictions, self-citation load-bearing steps, or ansatz smuggling is present. The abstract and described comparisons treat the techniques as external tools applied to independent experimental data; efficiency claims rest on empirical comparison rather than any reduction to the paper's own inputs by construction. This is the expected non-finding for an applied-methods instrumentation paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are mentioned in the abstract.

pith-pipeline@v0.9.0 · 5877 in / 1045 out tokens · 27490 ms · 2026-05-25T15:25:46.125614+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Radio detection of Cosmic-Ray Air Showers and High-Energy Neutrinos,

    Schr¨ oder, Frank G., “Radio detection of Cosmic-Ray Air Showers and High-Energy Neutrinos,” Prog. Part. Nucl. Phys. , vol. 93, pp. 1–68, 2017

    work page 2017

  2. [2]

    Measurement of cosmic-ray air showers with the Tunka Radio Extension (Tunka-Rex),

    P. A. Bezyazeekov et al. , “Measurement of cosmic-ray air showers with the Tunka Radio Extension (Tunka-Rex),” Nucl. Instrum. Meth. , vol. A802, pp. 89–96, 2015

    work page 2015

  3. [3]

    The TAIGA experiment: From cosmic-ray to gamma-ray as- tronomy in the Tunka valley,

    N. Budnev et al. , “The TAIGA experiment: From cosmic-ray to gamma-ray as- tronomy in the Tunka valley,”Nucl. Instrum. Meth. , vol. A845, pp. 330–333, 2017

    work page 2017

  4. [4]

    Tunka Advanced Instrument for cosmic rays and Gamma As- tronomy,

    D. Kostunin et al. , “Tunka Advanced Instrument for cosmic rays and Gamma As- tronomy,” in 18th International Baikal Summer School on Physics of Elementary Particles and Astrophysics: Exploring the Universe through multiple messengers (ISAPP-Baikal 2018) Bolshie Koty, Lake Baikal, Russia, July 12-21, 2018 , 2019

    work page 2018

  5. [5]

    Radio measurements of the energy and the depth of the shower maximum of cosmic-ray air showers by Tunka-Rex,

    P. A. Bezyazeekov et al. , “Radio measurements of the energy and the depth of the shower maximum of cosmic-ray air showers by Tunka-Rex,” JCAP, vol. 1601, no. 01, p. 052, 2016

    work page 2016

  6. [6]

    Reconstruction of cosmic ray air showers with Tunka- Rex data using template fitting of radio pulses,

    P. A. Bezyazeekov et al. , “Reconstruction of cosmic ray air showers with Tunka- Rex data using template fitting of radio pulses,” Phys. Rev. , vol. D97, no. 12, p. 122004, 2018

    work page 2018

  7. [7]

    Simulating radio emission from air showers with CoREAS,

    T. Huege, M. Ludwig, and C. James, “Simulating radio emission from air showers with CoREAS,” AIP Conf.Proc., vol. 1535, p. 128, 2013

    work page 2013

  8. [8]

    Representation learning: A review and new perspectives,

    Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 35, pp. 1798– 1828, Aug. 2013

    work page 2013

  9. [9]

    Chollet, “keras.” https://github.com/fchollet/keras, 2015

    F. Chollet, “keras.” https://github.com/fchollet/keras, 2015

    work page 2015

  10. [10]

    TensorFlow: Large-scale machine learning on heterogeneous sys- tems,

    M. Abadi et al. , “TensorFlow: Large-scale machine learning on heterogeneous sys- tems,” 2015. Software available from tensorflow.org

    work page 2015

  11. [11]

    Advanced functionality for radio analysis in the Offline soft- ware framework of the Pierre Auger Observatory,

    P. Abreu et al. , “Advanced functionality for radio analysis in the Offline soft- ware framework of the Pierre Auger Observatory,” Nucl.Instrum.Meth., vol. A635, pp. 92–102, 2011

    work page 2011

  12. [12]

    Imporoving reconstrucion methods for radio measure- ments with Tunka-Rex,

    P. A. Bezyazeekov et al. , “Imporoving reconstrucion methods for radio measure- ments with Tunka-Rex,” in 25th European Cosmic Ray Symposium (ECRS 2016) Turin, Italy, September 04-09, 2016, eConf C16-09-04.3 , 2017

    work page 2016

  13. [13]

    Improved measurements of the energy and shower maximum of cosmic rays with Tunka-Rex,

    D. Kostunin et al. , “Improved measurements of the energy and shower maximum of cosmic rays with Tunka-Rex,” PoS, vol. ICRC2017, p. 400, 2017

    work page 2017

  14. [14]

    Classification and Recovery of Radio Sig- nals from Cosmic Ray Induced Air Showers with Deep Learning,

    M. Erdmann, F. Schlter, and R. Smida, “Classification and Recovery of Radio Sig- nals from Cosmic Ray Induced Air Showers with Deep Learning,” JINST, vol. 14, no. 04, p. P04005, 2019

    work page 2019

  15. [15]

    Search for EAS radio-emission at the Tien-Shan shower instal- lation at a height of 3340 m above sea level,

    A. Beisenova et al., “Search for EAS radio-emission at the Tien-Shan shower instal- lation at a height of 3340 m above sea level,” EPJ Web Conf. , vol. 145, p. 11003, 2017

    work page 2017