Detecting FRB by DANCE: a method based on DEnsity ANalysis and Cluster Extraction

TL;DR

DANCE addresses the challenge of detecting weak, narrow-band FRBs by operating directly on the original time–frequency spectrum and using density-based clustering to identify localized high-density regions corresponding to FRB emission. The method combines 2-D DWT-based RFI mitigation, binary signal density space construction, and DBSCAN clustering, followed by Z-score and pulse-width-based cluster extraction to isolate FRB candidates. Simulations show high recall and precision for SNRs above 5, while real data from FRB20201124A reveal many weak bursts near stronger events, illustrating practical gains and the value of visual inspection. Overall, DANCE provides a robust, unsupervised pre-screening tool that can complement dedispersion-based searches and accelerate discovery of weak FRBs in large datasets.

Abstract

Fast radio bursts (FRBs) are transient signals exhibiting diverse strengths and emission bandwidths. Traditional single-pulse search techniques are widely employed for FRB detection; yet weak, narrow-band bursts often remain undetectable due to low signal-to-noise ratios (SNR) in integrated profiles. We developed DANCE, a detection tool based on cluster analysis of the original spectrum. It is specifically designed to detect and isolate weak, narrow-band FRBs, providing direct visual identification of their emission properties. This method performs density clustering on reconstructed, RFI-cleaned observational data, enabling the extraction of targeted clusters in time-frequency domain that correspond to the genuine FRB emission range. Our simulations show that DANCE successfully extracts all true signals with SNR~>5 and achieves a detection precision exceeding 93%. Furthermore, through the practical detection of FRB 20201124A, DANCE has demonstrated a significant advantage in finding previously undetectable weak bursts, particularly those with distinct narrow-band features or occurring in proximity to stronger bursts.

Figures (12)

• Figure 1: The detection process.
• Figure 2: RFI mitigation using the 2-D DWT for strong FRB signals. (a) The original data containing three FRB bursts, with the top panel showing the integrated signal. (b) The wavelet decomposition at the third level, displaying the four subcomponents "LL, LH, HL, and HH", respectively. Among these, "LL" represents the low-frequency approximation of the raw data, "HL, LH" extract horizontal and vertical features of the data, respectively, and "HH" captures fine-scale details. The decomposition is performed over five levels, with the third-level components shown here as representative examples. (c) RFI identification, where blanked regions indicate flagged interference at a 1 $\sigma$ threshold. "HL" component is retained unflagged, as the desired FRB signals are primarily represented in this channel. (d)The reconstructed data obtained after RFI removal from the decomposed components. All panels share an identical color scale for direct comparison.
• Figure 3: Transformation from the reconstructed spectrum to the signal density space (SDS). The left panel shows the reconstructed data; the middle panel displays the smoothed data; and the right panel presents the corresponding signal density representation. A weak FRB signal is visible within the region at 0.04-0.045 s in time and approximately 1300-1500 MHz in frequency.
• Figure 4: DBSCAN-identified clusters. The top color bar indicates different clusters. Left: all separated cluster samples in the SDS of Fig.\ref{['fig_density']}, with each color representing a distinct cluster. Middle: the size of each cluster sample, measured by the proportion of pixels each cluster occupies. Two notable clusters are labeled -1 and 13180 in the color bar, with sizes of 54% ($\log_{10} = -0.64$) and 1% ($\log_{10} = -2$), respectively. Right: the five largest cluster samples, including the background noise (label -1), an FRB signal (label 13180), and three other noise clusters.
• Figure 5: Cluster extraction based on Z-score and pulse width. The color bars at the top indicate the cluster labels. Left: Z-score of all cluster sizes in the SDS of Fig.\ref{['fig_density']}, each color indicates a cluster. Middle: the pulse width of each cluster sample. Right: the filtered FRB cluster with a Z-score of 3.6 and a pulse width of 4.1 ms, respectively. This cluster has been filtered out based on a criteria of $Z>3$ and $W<100$.