PaperlayerView exact revision on arXiv

Source-linked AI summary

Outage-Aware Robust Sensing for UAV ISAC Systems

Qiming Li, Lei Zhang, Haoran Xu, Lina Mohjazi

arXiv:2610.11720v1eess.SPcs.ITmath.IT

TL;DR

Dynamic blockage makes the informative sensing links at the next update uncertain, while a single predicted pattern may miss low-information outcomes and protecting all patterns may waste power or become infeasible. The paper uses a history-aware joint predictor and conformal pattern sets to allocate power robustly over serviceable outcomes while treating all-blocked outcomes as physical outages. In simulation, the serviceability-conditioned policy uses less average power than protection over all serviceable patterns.

  • Problem

    Dynamic blockage makes next-slot informative sensing links uncertain, creating a need to account for correlated availability patterns before the pattern is revealed.

    Source: §I · p. 1

  • Method

    A history-aware joint predictor and adaptive conformal sets retain plausible correlated patterns, with robust tracking constraints applied only to serviceable patterns and all-blocked outcomes handled as outages.

    Source: §I · p. 1

  • Results

    Avg. power: 20.23±0.37 for Conformal-serviceable versus 38.14±0.35 for All-serviceable.

    Source: B. Prediction and Calibration · p. 5

  • Takeaways & Limitations

    The serviceability split avoids both under-protection of uncertain serviceable outcomes and enforcing an all-blocked state that power cannot repair.

    Source: C. Main Policy Comparison · p. 6

Abstract

from arXiv · show

Integrated sensing and communications (ISAC) infrastructure can provide an external perception layer for aerial robots, but dynamic blockage makes the set of informative sensing links at the next update uncertain. A single predicted availability pattern can miss low-information outcomes, whereas protecting all patterns wastes power and may become infeasible when all links are blocked. We propose an outage-aware conformal robust power-allocation framework for multi-BS UAV tracking. A history-aware joint predictor maps radar and tracking histories to next-slot availability probabilities, and adaptive prediction sets retain plausible correlated patterns. Tracking constraints are enforced only for retained non-outage patterns, while the all-blocked pattern is treated separately as a physical sensing outage. With per-slot beam directions fixed, the resulting scalar-power allocation is a semidefinite program. A reliability bound relates serviceable-slot tracking failure to deployed-policy miscoverage and optimization infeasibility. Across paired closed-loop simulations over five random seeds, the method attains a 0.01% serviceable-slot constraint-violation rate and uses 47.0% less power than protection over all non-outage patterns.

I. INTRODUCTION

Dynamic blockage makes informative next-slot sensing links uncertain, and decisions must account for correlated patterns and the closed-loop effect of sensing actions on later observations. The paper combines joint availability prediction with serviceability-conditioned robust allocation for UAV tracking.

I. INTRODUCTION

  • Dynamic blockage and common blockers make next-slot sensing-link availability uncertain, while the current power choice affects next-slot observation quality.The allocator acts before the next joint pattern is revealed, and the sensing action changes observations later used by the tracker and predictor.Source: §I · p. 1
  • The proposed framework retains plausible correlated patterns, enforces tracking constraints on serviceable patterns, and treats all-blocked outcomes as physical outages.With per-slot beam directions fixed, scalar-power allocation becomes an SDP, and a coverage-to-tracking bound separates miscoverage, infeasibility, and outage.Source: §I · p. 1

A. Network and Tracking Model

  • The model considers multiple sensing-capable base stations tracking a UAV that does not receive communication service.It uses a nearly constant velocity state model and an extended Kalman filter for tracking.Source: A. Network and Tracking Model · pp. 1–2

B. Sensing-Link Availability

Sensing-link availability is a binary label for whether the environment permits an effective radar update under the admissible power budget. The joint availability vector represents the next sensing pattern across base stations.

B. Sensing-Link Availability

  • A link is labeled available when its maximum-admissible-power SCNR meets the minimum detection threshold.This makes availability an environment-permitted sensing capability rather than a judgment about whether the current allocation illuminates the link sufficiently.Source: B. Sensing-Link Availability · p. 2
  • The joint availability vector comprises binary link labels and indexes one of 2^B possible patterns.Each pattern corresponds to a particular combination of available and unavailable base-station links.Source: B. Sensing-Link Availability · p. 2

C. Serviceable Patterns and Sensing Outage

The paper separates patterns with potentially informative links from the all-blocked sensing outage. Its propagation model represents link blockage and the radar and communication channels used in the system.

C. Serviceable Patterns and Sensing Outage

  • Serviceable patterns contain at least one potentially informative link, whereas the all-blocked pattern cannot produce a valid radar update under the admissible power budget.The all-blocked pattern is treated as a sensing outage rather than a beamforming failure.Source: C. Serviceable Patterns and Sensing… · p. 2

D. Propagation, Blockage, and Channel Models

  • Each sensing link has a line-of-sight, partial-blockage, or complete-blockage state, and blockage may be temporally and cross-link correlated.The associated attenuation decreases from line-of-sight through partial blockage to complete blockage.Source: D. Propagation, Blockage, and Channel… · p. 2
  • The maximum SCNR is based on the radar equation and depends on sensing gains, wavelength, radar cross section, receiver noise, and residual clutter.The maximum admissible probing power is among the quantities used in this link model.Source: D. Propagation, Blockage, and Channel… · p. 2
  • The downlink channel model includes distance-dependent path loss, shadowing, blockage attenuation, and scattered components, with full network CSI available to the allocator.The channel description includes a deterministic line-of-sight array response and spatial covariance for the scattered component.Source: D. Propagation, Blockage, and Channel… · pp. 2–3

E. Radar Observations and Closed-Loop Timing

Radar measurements update the tracker only for available links, and their precision depends on the preceding target-directed illumination. Echo power and detection flags follow the same timing.

E. Radar Observations and Closed-Loop Timing

  • An unavailable link contributes no valid measurement update, while an available link provides a radar measurement to the EKF.The EKF evaluates the measurement Jacobian at the predicted state.Source: E. Radar Observations and Closed-Loop… · p. 3
  • For an available link, inverse measurement covariance increases with preceding-slot illumination, which scales unit-illumination precision.The precision coefficients correspond to range, azimuth, and radial velocity.Source: E. Radar Observations and Closed-Loop… · p. 3
  • Observed echo power and the binary detection flag follow the same timing, with detection modeled from echo and a fixed false-alarm rate under blockage.Source: E. Radar Observations and Closed-Loop… · p. 3

A. History-Aware Joint Pattern Predictor

The predictor uses radar, tracking, and motion histories to produce a ranked joint distribution over next-slot availability patterns. A shared recurrent encoder and joint output preserve cross-link dependence.

A. History-Aware Joint Pattern Predictor

  • The predictor maps histories to a joint probability vector over next-slot patterns because common blockers correlate sensing links.Its output is a ranked joint distribution rather than a hard link decision.Source: A. History-Aware Joint Pattern Predictor · p. 3
  • Radar and tracking features combine detection and echo evidence, relative sensing geometry, tracking innovations, prior illumination, and consecutive-miss counts.Shared motion context adds estimated velocity-derived speed and heading; covariance eigenvalues encode tracking confidence.Source: A. History-Aware Joint Pattern Predictor · p. 3
  • A shared gated recurrent unit encodes each length-L history, with larger L providing more temporal context.The shared weights capture common dynamics across the histories.Source: A. History-Aware Joint Pattern Predictor · p. 3
  • The joint output estimates each pattern’s next-slot probability, and training minimizes negative log-likelihood because APS uses its ranking.The output combines per-BS representations with shared context.Source: A. History-Aware Joint Pattern Predictor · p. 3

B. Adaptive Prediction Set Calibration

Adaptive prediction sets retain plausible next-slot patterns, while calibration is based on held-out episodes under a fixed reference policy. Fixed per-slot directions keep the power allocation tractable and allow sensing handover through power redistribution.

B. Adaptive Prediction Set Calibration

  • APS forms a variable-size set by including probability-ranked patterns through the first whose cumulative probability reaches the calibrated threshold.Calibration uses held-out episodes under a fixed All-serviceable reference policy; its stated guarantee assumes exchangeability between calibration and test scores generated by that policy.Source: B. Adaptive Prediction Set Calibration · p. 3
  • Deployed actions can shift histories, making prediction-set coverage policy-dependent.The calibration passage specifies a fixed reference policy.Source: B. Adaptive Prediction Set Calibration · p. 3
  • The serviceable-set size determines the robust constraint count, while a separate flag records a plausible power-insensitive outage; an empty serviceable set yields communication-only allocation.The all-blocked state cannot be improved by power.Source: B. Adaptive Prediction Set Calibration · p. 3

A. Fixed-Direction Transmission

  • Fixing directions within each slot avoids joint beam–power optimization and keeps the FIM affine in scalar powers.Sensing handover is achieved by setting a sensing power to zero and redistributing power without losing convexity.Source: A. Fixed-Direction Transmission · pp. 3–4

B. Communication Constraints

The scalar power vector includes communication and sensing powers, subject to user quality-of-service targets and per-BS power budgets. With fixed directions, the communication constraint and local budgets are linear.

B. Communication Constraints

  • The allocation vector combines communication-user powers and sensing powers, and must satisfy communication QoS targets and per-BS power budgets.The passages define the communication and sensing power components before specifying fixed-direction gains.Source: B. Communication Constraints · p. 4
  • With fixed directions, the communication constraint and local power budgets are linear.Source: B. Communication Constraints · p. 4
  • Larger user SINR targets improve communication protection but leave less power for sensing.Source: B. Communication Constraints · p. 4

C. Pattern-Dependent Tracking Information

Each availability pattern changes the sensing geometry, so the same power allocation is evaluated separately for each realization. Tracking constraints use the resulting pattern-dependent sensing information, while the all-blocked pattern is treated as outage.

C. Pattern-Dependent Tracking Information

D. Serviceability-Conditioned Robust Power Allocation

The allocator minimizes total power subject to communication and tracking constraints for every retained serviceable pattern, with fixed directions enabling an SDP formulation. A coverage-to-tracking bound separates serviceable-slot failures due to miscoverage or infeasibility, while accounting separately for outage.

D. Serviceability-Conditioned Robust Power Allocation

  • With fixed per-slot directions, minimizing total power under communication and tracking constraints for every retained serviceable pattern admits an SDP formulation.The SDP uses one Schur-complement LMI per retained serviceable pattern.Source: D. Serviceability-Conditioned Robust… · p. 4
  • For a fixed deployed policy, serviceable-slot tracking violations are linked to prediction-set miscoverage and robust-allocation infeasibility; physical outage is accounted for separately.Under calibration–deployment exchangeability, miscoverage is bounded by α; feasibility on every serviceable slot makes the infeasibility probability zero.Source: D. Serviceability-Conditioned Robust… · pp. 4–5
  • Each slot updates the EKF, predicts availability, constructs an APS set, and separates its serviceable patterns before allocation.The control loop uses observations generated by the previous transmit action.Source: D. Serviceability-Conditioned Robust… · p. 5
  • The controller forms communication and sensing directions before solving the robust allocation, and uses a worst-trace fallback if that solve is infeasible.The chosen transmit action affects radar observations at the next slot.Source: D. Serviceability-Conditioned Robust… · p. 5

V. SIMULATION RESULTS

In five-seed closed-loop simulations, the conformal-serviceable policy used less average power than the All-serviceable policy, with similar reported Aall values. APS calibration retained a compact set of patterns and had high empirical coverage in the tested regime.

A. Setup and Metrics

  • Results average five seeds in a 1000 m×1000 m scene with three ISAC base stations, two communication users, and a UAV at 120 m altitude.Each seed used 300 training, 60 validation, 60 calibration, and 50 test episodes, with 60 slots per episode; performance includes serviceable-slot violation and outage-inclusive Aall.Source: A. Setup and Metrics · p. 5

B. Prediction and Calibration

  • On held-out validation episodes, the GRU achieved 98.4% top-2 accuracy, while APS retained 3.0 patterns on average and reached 99.96% empirical coverage at a nominal 90% level.Coverage remained above 99.3% under deployed policies.Source: B. Prediction and Calibration · p. 5
  • 20.23±0.37 versus 38.14±0.35 dimensionless units of average power for Conformal-serviceable and All-serviceable, respectively, while their Aall values were 5.74±0.71% and 5.73±0.71%.Table I reports means ± SEM over five seeds; Fig. 2 plots power and reliability, with the lower-left region preferred.Source: B. Prediction and Calibration · p. 5, Table I · p. 5, Figure 2 · p. 5

C. Main Policy Comparison

The policy protects plausible serviceable patterns while treating all-blocked states as physical outages. Results show lower power than All-serviceable protection, low serviceable-slot violation across tested shifts, and increasing power savings as the number of base stations grows.

C. Main Policy Comparison

  • 47.0% less power is used than with All-serviceable protection by pruning implausible patterns while retaining robust protection over plausible serviceable patterns.Serviceability conditioning places the method close to the outage floor.Source: C. Main Policy Comparison · p. 5

D. Risk, Shift, and Scalability

  • At B = 6, an average of 7.53 constraints are retained for 63 serviceable patterns; increasing B lowers power and enlarges savings over All-serviceable, with modest violation change.For larger networks, 2B predictor outputs and one LMI per retained pattern can increase computation and memory costs.Source: Figure 3 · p. 6, D. Risk, Shift, and Scalability · p. 6
  • Below 0.1% violation is achieved by Conformal-serviceable across six shifts, versus up to 23.86% for MAP, with the nominal predictor and APS threshold frozen.The shifts include BS layout, denser or faster blockers, UAV turns, higher detection thresholds, and noisier SCNR.Source: D. Risk, Shift, and Scalability · p. 6

VI. CONCLUSION

  • The framework calibrates joint availability prediction sets, separates serviceable patterns from physical sensing outage, and formulates fixed-direction scalar-power allocation as an SDP.Its reliability analysis attributes serviceable-slot failures to prediction-set miscoverage or optimization infeasibility, while accounting separately for physical outage.Source: §VI · p. 6

Related papers

Loading 2610.11720v1…