使用監督式機器學習以 TAROGE 搜尋快速電波爆發

Abstract

本研究針對快速電波爆發（FRB）的探測設計了一套以監督式機器學習為基礎的分析流程，並以 TAROGE-3 所蒐集的資料訓練模型並嘗試搜索相關訊號。TAROGE-3 原本是針對宇宙射線探測所設計的無線電波測站，為了可能的 FRB 探測，在 2021 到 2022 年間，曾以 20 赫茲固定頻率的強制觸發 (forced trigger) 採集最高至 350 MHz 、長度 1.2 微秒的訊號。本研究以殘差神經網路為架構打造客製化的機器學習模型，以多頻道的電壓振幅時頻譜為輸入資料，判斷是否存在類似 FRB 的反啁啾訊號。完成訓練的模型可以剔除約 99.8% 的背景資料，並在探測訊號方面維持高度效率。我們接著運用顯著圖分析以及訊號的方位重建以訂定另一個篩選標準，這使得最終的偽陽性率得以再下降約一個數量級。 本論文將介紹 FRB 的相關特性、TAROGE-3 系統，及監督機器學習的基本概念，並討論本研究如何處理資料、訓練模型、驗證結果和設定參數，最後呈現訊號搜尋的結果。

In this study, we developed an analysis process based on supervised machine learning for the detection of Fast Radio Bursts (FRBs). The model was trained using data collected by TAROGE-3, originally designed as a radio telescope for cosmic ray detection. To facilitate FRB detection, a periodic forced trigger was configured to record 1.2 µs long waveforms with 350 MHz analog bandwidth, and data were collected at a rate of 20 Hz between 2021 and 2022. A custom machine learning model based on a residual neural network was developed, using multi-channel voltage amplitude spectrograms as input data to determine the presence of anti-chirp signals similar to FRBs. The trained model can filter out approximately 99.8% of the background data while maintaining high efficiency in signal detection. We then applied saliency map analysis and signal direction reconstruction, which further reduced the fake rate (false positive rate) by about an order of magnitude. This thesis introduces the characteristics of FRBs, the TAROGE-3 system, and the basic concepts of supervised machine learning. It discusses data processing, model training, result validation, and parameter setting. Finally, the results of the signal search are presented.