Why We Kept Multistatic Fusion Off the Critical Path in Our ISAC Sensing Plane
The decision under review is where sensing truth lives when the network is partially broken. We chose per-site monostatic sensing as the primary source of tracks, with multistatic fusion across gNBs added as a layer the system can drop without anyone noticing at the service level. The rejected option was a fusion-first architecture, where bistatic and multistatic measurements from coordinated cells feed a central sensing function that owns every track. This record states that option as its proponents would, explains why it lost, and names the condition that would have reversed the call.
The option we rejected, stated fairly
Fusion-first is the stronger design when everything works. A central sensing function that ingests bistatic echoes from several transmit and receive pairs gets geometric diversity that no single site can match: better angular coverage, resolution of targets that sit along a monostatic sector's blind radial, more resistance to clutter and multipath, and a single track identity across cell borders without handover logic. It also avoids monostatic's hardest hardware problem, which is full-duplex self-interference cancellation at the radio unit, because the transmitter and receiver sit at different sites. The study work in 3GPP, from the Rel 19 ISAC channel modelling through the early 6G study items, treats bistatic and multistatic geometries as first-class deployment cases for good reason. If we were designing for peak detection performance in a fully healthy network, this option wins.
Where it breaks: timing is a shared dependency
Bistatic range comes from time of flight across two sites, so any timing offset between transmitter and receiver turns directly into range bias, at the speed of light: a synchronisation error of 10 ns puts roughly 3 m of bias into the bistatic range sum. That is the parameter that decided this record. Under normal operation, PTP over the transport network plus GNSS discipline keeps the error well inside a range cell. The failure mode we could not accept is a GNSS outage or PTP path failure that pushes sites into holdover. The radios keep transmitting, the communication link stays up because its timing tolerance is far looser, and the fused tracks keep coming out of the central function with plausible covariance. They are wrong, and nothing in the fusion output says so. A degraded comms link announces itself through HARQ statistics and throughput. A biased multistatic track looks healthy until it disagrees with ground truth, and in a vehicle or drone safety use case that disagreement is found by the consumer of the track, not by us.
Monostatic sensing carries its own timing reference inside a single radio unit, so a site losing external sync loses fusion quality, not its own range accuracy, and that isolation is the whole point of the chosen design.
The chosen design and the condition that would reverse it
Each site produces tracks from its own echoes and publishes them with a sync health flag taken from the timing subsystem. The fusion layer consumes only measurements from pairs whose flags show both ends locked, and when a pair drops into holdover the layer removes it and falls back to associating monostatic tracks, which degrades accuracy at cell edges but never injects bias. Our decision rule is this: multistatic fusion may become the primary track source only where the worst-case holdover drift over the longest expected sync outage stays below one range resolution cell, as measured and not as specified by the oscillator datasheet. By our estimate, that holds today only at sites with high-stability oscillators and redundant timing paths, which is a minority of the planned footprint. The price is real: we pay for full-duplex cancellation at every sensing site, and we give up coverage of targets on monostatic blind radials until fusion is healthy again.
Technical FAQ: Why not keep fusion-first and just widen the track covariance when a site enters holdover?
Because holdover drift is a bias, not noise, and inflating covariance around a biased mean still points downstream consumers at the wrong place with false confidence. Widening the uncertainty enough to cover the drift makes the fused track worse than the monostatic one it replaced, so excluding the pair is both simpler and more accurate. The question we pass to the next design stage is whether downstream consumers will accept a track source that changes its accuracy class during a sync event, or whether that change has to be exposed explicitly in the sensing service API.