RF-VoID: Towards Bandwidth-Efficient Exterior Tile Void Detection via Narrowband Radio-Frequency Representation Learning

• Hidden debonding can form as tile-layer bonds weaken, creating a falling-tile hazard that motivates periodic inspection.Sounding and thermography have limitations including labor and access demands, inspector and environmental sensitivity, and dependence on thermal contrast.Source: §1 · p. 1
• Conventional range-domain imaging becomes bandwidth-dependent when void thickness approaches or falls below effective depth resolution, while wide bandwidth is costly to deploy.Bandwidth also determines acquisition time and data volume, and constrains regulated emissions and pole- or drone-mounted payloads.Source: §1 · p. 2
• Reconstructing an image before classification preserves the imaging resolution limit, while sparse recovery and deconvolution depend on accurate system estimates and stable conditions.Those conditions are not guaranteed on weathered facades.Source: §1 · p. 2
• A void perturbs amplitude and phase across the measured band, so direct sub-band decisions need not form a range peak to detect it.The encoder and objective address frequency ordering, specimen-varying layer geometry, point-wise decisions, class imbalance, and the higher cost of missed voids.Source: §1 · p. 2
• At 0.5 GHz, RF-VoID reaches 98.61% accuracy and 95.84% overall F1-score under the mixed-sample protocol.The authors present this as evidence that inspection bandwidth can be traded for computation.Source: §1 · p. 2

2.1 Microwave and Millimeter-Wave Inspection of Layered Civil Structures

• Wideband millimeter-wave measurements have visualized tile voids, but conventional image-based detection is bounded by reconstruction quality and bandwidth.Reported detectability in this body of work uses laboratory-grade wideband instrumentation, while limited field bandwidth is not addressed.Source: §2.1 · p. 3

2.2 Learning-Based Interpretation of Radar Measurements

• Learning-based radar inspection commonly classifies formed images, so it cannot resolve defects that the physical reconstruction leaves unresolved.The same image-first pattern appears in GPR and terahertz inspection examples.Source: §2.2 · p. 3

2.3 Learning Directly from Complex Radio-Frequency Responses

• Direct signal learning has been demonstrated for layered composites, but prior work requires near-field access and leaves bandwidth reduction unexamined.RF sensing studies show learnable signatures in complex responses under more favorable conditions, unlike facade inspection's buried voids, varying backgrounds, and point-wise scan decisions.Source: §2.2 · p. 3, §2.3 · p. 4

3.1 Millimeter-Wave Measurement System

• A one-port vector network analyzer, broadband horn, Teflon lens, and 3-D stage form the stepped-frequency measurement system.The lens focuses the field on a localized wall region; the real-aperture arrangement reduces acquisition and processing complexity relative to SAR and array systems.Source: §3.1 · pp. 4–5
• The system records 2048 frequency points from 4–40 GHz at each spatial position, with amplitude and phase variations providing a basis for void detection.The complex I/Q response can retain defect-sensitive information even when a narrowband signal does not produce a clearly separated range profile.Source: §3.1 · p. 4

3.2 Exterior-Wall Samples

• Specimens reproduce tile, adhesive mortar, base mortar, and concrete layers, with artificial voids measuring 50 × 50 mm or 100 × 100 mm laterally.Thin inserts create controlled void regions after curing.Source: §3.2 · p. 5
• M-series voids are embedded in base mortar, while T-series voids occupy several interfaces, including separated interfaces.The configurations vary void depth, thickness, interface location, and structural complexity.Source: §3.2 · p. 5
• Measurements use a 0.5 cm spatial grid and are represented as 60 × 60 × 2048 × 2 arrays, with reduced-bandwidth subsets drawn from the wideband data.The array dimensions encode spatial position, frequency sample, and I/Q channels.Source: §3.2 · pp. 5–6

4.1 Thin-Void Detection Using A-Scan Signals

• At 36 GHz, physical depth resolution is approximately 1.9–3.0 mm, exceeding the specimens’ 0.5–1.0 mm void thicknesses.Thus, even the full measurement bandwidth should not produce two clearly separated boundary peaks for a thin void.Source: §4.1 · p. 6
• At 4–26 GHz, T4’s A-scan resembles the full-band result, whereas 24–26 GHz produces pronounced ringing and peak broadening.The M4 perturbation is weaker and occurs farther in range than T4’s; the figure marks void responses at approximately 10 mm and 25 mm depths.Source: §4.1 · p. 6, Figure 3 · p. 7

4.2 Imaging-Based Inspection and Its Limitations

• At 4–26 GHz, B-scans still clearly recognize void regions, while narrower-band signatures broaden and become less separable along depth.The 4–18 GHz case lies near the lower limit for visual identification; deconvolution-based restoration depends strongly on parameter selection, wavelet estimation, signal-to-noise ratio, and sample conditions.Source: §4.2 · p. 7, Figure 4 · p. 8
• C-scans show wider-band void regions more recognizably, but narrowband slices can leave deeper M4 voids weak and obscure T7 responses.The representative depth slices were selected from void-related responses seen in A- and B-scans.Source: §4.2 · p. 7, Figure 5 · p. 8
• C-scan continuity can help distinguish void anomalies from local clutter, but it requires dense two-dimensional acquisition and careful depth selection.A slightly mismatched extraction depth can substantially weaken void contrast, while array- and SAR-based imaging carry higher hardware, calibration, data, and computational burdens than simpler narrowband sensing.Source: §4.2 · p. 8

5.1 From Range-Domain Imaging to Learned RF Representation

• A void modulates amplitude and phase across the measured sub-band, so RF-VoID discriminates its spectral signature instead of reconstructing a range-domain peak.When bandwidth is too small to separate void boundaries from neighboring interfaces, the modulation remains spread across the band and superimposed on stronger responses.Source: §5.1 · pp. 9–10

5.2 Physics-Informed Narrowband RF Representation

• The input retains ordered in-phase and quadrature samples alongside magnitude and phase, preserving complex response patterns without range-profile conversion.Magnitude and phase provide explicit cues about reflection strength and relative propagation delay.Source: §5.2 · p. 10

5.3 Dual-Branch Narrowband RF Encoder

• The encoder combines local frequency-axis convolutions with locality-biased attention, using relative frequency separation to capture both short-scale fluctuations and band-wide trends.Relative encoding avoids tying patterns to absolute frequency indices, while head-specific distance penalties inspect multiple effective frequency widths.Source: §5.3 · pp. 10–11

5.4 Inspection-Oriented Training Objective

• The training objective combines focal loss for class imbalance, an auxiliary penalty on labeled voids to favor recall, and entropy regularization for decisive predictions.Normal positions outnumber void positions by roughly an order of magnitude, and missed voids are treated as more costly than false alarms.Source: §5.4 · p. 12

5.5 Defect Probability Map Generation

• The trained model writes a whole-sub-band void posterior at each scan position, avoiding depth-slice selection, range-profile reconstruction, deconvolution, and source-wavelet estimation.The point-wise mapping inherits the scan's spatial sampling and consumes only the selected sub-band bandwidth.Source: §5.5 · p. 12

6.1 Experimental Protocols

• The mixed-sample split uses data from all specimens in training, validation, and testing, whereas leave-one-specimen-out testing holds one of twelve specimens entirely out.Mixed-sample results average three training runs; leave-one-specimen-out results aggregate twelve folds. At 0.5 GHz, the dataset has about 9.7 normal positions per void position.Source: §6.1 · p. 13

6.2 Mixed-Sample Detection Performance

• At 1 GHz, the model retains a 97.44% macro F1-score and 95.36% void-class F1, supporting the hypothesis that thin voids perturb frequency-domain response despite poor conventional B-scan or C-scan separability.Source: §6.2 · p. 14
• Figure 8 maps show cleaner, more spatially consistent void predictions at 16 and 8 GHz than at narrower bandwidths.Maps become noisier and boundaries less stable as bandwidth decreases, but many main void locations remain identifiable at 0.5 GHz.Source: Figure 8 · p. 14, §6.2 · p. 14

6.3 Leave-One-Specimen-Out Generalization Performance

• RF-VoID achieves the highest mean macro F1-score at 0.5, 1, and 2 GHz among the compared models in leave-one-specimen-out testing.Figure 9 summarizes mean macro F1 across twelve held-out-specimen folds, with standard-deviation and min–max ranges.Source: §6.3 · p. 14, Figure 9 · p. 15
• A 6.45-point mean macro F1 decrease occurs from 1 GHz to 0.5 GHz, while 2 GHz and 1 GHz show no detectable change.The 0.5-GHz decrease is significant (W = 3, p = 0.002); the 2-to-1-GHz comparison is not (W = 28, p = 0.42).Source: §6.3 · p. 15

6.4 Ablation Study

• At 0.5 GHz, the full complex RF input consistently outperforms in-phase-only, quadrature-only, and raw I/Q variants.The ablation supports combining amplitude, phase, and raw complex frequency-domain information.Source: Table 4 · p. 16, §6.4 · p. 16

7 Conclusion

• 98.61% accuracy and 95.84% overall F1 are achieved at 0.5 GHz, a seventy-two-fold bandwidth reduction from the full sweep, without range-profile reconstruction, deconvolution, or depth-slice selection.Conventional A-, B-, and C-scan analysis finds 0.5–1.0 mm voids unresolved even at the 4–40 GHz sweep.Source: §7 · p. 16
• RF-VoID processes measured sub-band responses in frequency order, combining amplitude and phase with in-phase and quadrature channels in a dual-branch encoder.The attention branch uses relative position encoding and a distance-dependent locality bias; the objective accounts for class imbalance and the asymmetric cost of missed voids.Source: §7 · p. 16
• Generalization remains limited across walls differing in layer thickness, material, and surface condition, and performance drops significantly at 0.5 GHz on unseen specimens.The conclusion identifies broader specimen diversity as the next step toward field use.Source: §7 · p. 17