Real-Time RAN Observability at the Far Edge: AI on the Fronthaul

Abstract

Operational RAN observability is reported every 5 to 15 minutes, averaging over as many as 1.8 million slots. Resolution is only half the problem: the quantities that explain a physical-layer fault, such as the beamforming weights the DU computed and the elements each layer actually drove, are not counted at all. Neither limit will cope with 6G integrated sensing, non-terrestrial networks, and ever-larger MIMO arrays. The 7.2x fronthaul split puts a standardized, tappable boundary in the middle of every modern base station; we argue it is the right place to obtain real-time observability, and that the analysis must run at the far edge, because the evidence is transient: it crosses the wire once, cannot be averaged down without losing the fault, and is not recoverable afterwards at any resolution. From passive captures, with no vendor-supplied instrumentation and no cooperation from either the DU or the radio supplier, we recover DU-side scheduling behavior and radio-side execution, including the per-beam and per-layer beamforming weights carried on the C-plane, separating measured quantities from those conditioned on an array hypothesis. On that extraction we ground an eleven-agent platform in which a language model only interprets measurements. Across three independent capture labs covering 4T4R, 32T32R, and 64T64R radios, a fourth lab read over the M-plane, and a dataset of 87 distinct captures, one unmodified code path recovered per-layer weights and revealed three faults nobody was looking for, two of them on the DU side, including a pair of effectively un-beamformed layers for which no alarm currently exists. A separate root-cause diagnosis was confirmed by the repair that followed.