The Giant Radio Array for Neutrino Detection (GRAND) aims to detect
ultra-high-energy cosmic rays and neutrinos above $10^{17}$~eV with large-scale
arrays of autonomous radio detector units (DUs) operating in a self-trigger mode.
GRANDProto300 (GP300), the GRAND pathfinder array under deployment at the
Xiaodushan site near Dunhuang, China, is currently operating in its 65-DU stage
(GP65) and will expand to 300 DUs. Self-triggered radio arrays face two dominant
classes of background: accidental coincidences of independent random triggers,
dominated by Galactic noise, and repetitive emission from fixed anthropogenic
sources. We present a two-step offline cleaning pipeline developed for GP300 data:
(i) causality cleaning, which maps each event onto a graph of causally connected
DU pairs and extracts the event core with a maximum-clique algorithm, and (ii) a
time-difference (TD) fingerprint method that rejects events matching the
characteristic signatures of known background sources accumulated in a history
library. Applied to GP65 data, the causality cleaning reduces accidental events by
a factor of about 30, while the TD-fingerprint cut suppresses fixed-source events
by a factor of about 20, with a 92\% survival rate for cosmic-ray-like events in
toy Monte-Carlo validation. The pipeline delivers clean event samples for
cosmic-ray candidate searches and constitutes a key step towards fully autonomous
radio detection of air showers with GRAND.