{"id":"1f7dd268-c652-44c4-98cd-53ce314e8abc","arxiv_id":"2607.02545","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Transformer fusion of asynchronous PZT guided-wave and FBG strain data yields HI MAE/RMSE <0.1 and localization MAE/RMSE <0.0465/0.1571, beating single-sensor and SOTA DNN baselines by ~60% on ReMAP composite fatigue panels.","lead":"A Transformer fuses ultrasonic guided-wave (PZT) and fiber-optic strain (FBG) streams that arrive at different rates, jointly predicting a health indicator and a damage map on fatigued aircraft composite panels. The fused model cuts error roughly in half versus single-sensor or standard deep nets, offering a practical route to multi-modal structural health monitoring.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"AE-derived labels make reported fusion gains partly circular; independent damage ground truth is missing.","rationale":"The reader correctly isolates the AE-label assumption as the weakest link. The architecture, tokenization, cycle-aware encoding, and attention visualizations are carefully engineered and the leave-one-out numbers are clear; no mathematical contradiction appears. The concern is therefore not that the model fails to fit its targets, but that those targets are themselves derived from a third SHM modality whose completeness relative to true composite damage is unproven. That single issue keeps the paper at CONDITIONAL rather than ACCEPT and keeps correctness risk medium. An independent NDT or residual-property check would settle the circularity question without requiring new model development. No stronger load-bearing flaw (e.g., data leakage, incorrect loss formulation, or non-reproducible metrics) is evident from the text.","tokens_in":19815,"tokens_out":563,"duration_ms":6091,"concrete_test":"On at least one panel (preferably L04 and one in-distribution panel), obtain an independent damage map (ultrasonic C-scan or X-ray CT at several fatigue stages, or residual-stiffness / residual-strength curves). Recompute MAE/RMSE/SSIM of the Transformer predictions against this map instead of the AE-KDE heatmap; if the fusion advantage over PZT-only/FBG-only shrinks by more than ~30 % or R^{2} falls below 0.6, the reported gains are largely AE-circular.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (nearly 60 % HI improvement and best localization MAE/RMSE) rests on treating AE cumulative energy (averaged over four channels after Mo–Pr–Tr selection, §3.1) and Gaussian-KDE maps from AE TDoA triangulation (§3.2, Eqs. 17–24) as unbiased ground truth for both tasks. PZT and FBG never see independent labels: residual strength, stiffness drop, C-scan, or post-test NDT are never reported. Because AE is itself an SHM modality that can miss quiescent growth or be biased by sensor placement and anisotropic velocity assumptions (Eq. 18), the fusion model may simply be learning to reproduce AE-derived quantities rather than recovering true internal damage. The leave-one-out tables (Tables 3–4) therefore measure agreement with AE, not necessarily superior damage assessment. The OOD L04 case (longer life, edge impact) already shows elevated error, consistent with label-distribution shift rather than pure sensor-fusion failure. Without an AE-independent reference, the headline performance numbers remain circular with respect to the label source.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a Transformer-based early-fusion framework that tokenizes heterogeneous PZT ultrasonic guided-wave spectrograms and multi-channel FBG strain series, aligns them with cycle-aware temporal and modality embeddings plus binary padding masks, and jointly performs health-indicator (HI) regression and pixel-wise damage localization. Labels are derived from continuous AE: cumulative energy (selected via Mo–Pr–Tr) for HI and Gaussian-KDE maps from anisotropic TDoA triangulation for localization. On four ReMAP CFRP skin-stiffener panels under compression–compression fatigue, leave-one-panel-out results show fusion MAE/RMSE below ~0.08 for HI (claimed ~60% better than PZT-only, FBG-only, and 1D-CNN/Bi-LSTM/CNN-LSTM baselines) and localization MAE/RMSE below 0.0465/0.1571 with improved SSIM. Attention maps are used to illustrate cross-modal focus.","tokens_in":20161,"tokens_out":1565,"duration_ms":19497,"significance":"If the reported gains hold under stronger validation, the work is a useful contribution to aerospace SHM: it addresses a practical multi-rate fusion problem (PZT every 5000 cycles vs FBG every 500), provides an end-to-end multitask Transformer with explicit missing-modality handling, and supplies attention visualizations that aid interpretability. The experimental setting (real Embraer-design stiffened panels, fatigue to failure, multi-sensor ReMAP data) is more realistic than many synthetic SHM studies. Strengths include clear single-sensor and DNN baselines (Tables 3–5), quantitative localization metrics including SSIM, and an explicit reconstruction regularizer for sparse tokens. The main significance is methodological (heterogeneous token fusion under fatigue acquisition mismatch) rather than a definitive new physical damage metric.","major_comments":[{"comment":"§3.1–3.2 and Tables 3–4: All supervised targets are AE-derived (channel-averaged cumulative energy; Gaussian-KDE of TDoA sources under the anisotropic velocity model in Eqs. 17–24). PZT/FBG never see independent damage references (C-scan, residual stiffness/strength, post-test NDT, or even direct comparison of localization peaks to the known 10 J impact sites beyond the qualitative red boxes in Fig. 12). The headline “nearly 60% improvement” and “highest accuracy” therefore measure agreement with AE proxies, not verified internal damage. AE can miss quiescent growth and is sensitive to placement and velocity assumptions. Reframe claims as “prediction of AE-based HI/localization from PZT+FBG,” quantify agreement of AE maps with known impact locations, and discuss bias risk; ideally add at least one AE-independent check on a subset of cycles or panels.","section":"§3.1–3.2, Tables 3–4, Fig. 12"},{"comment":"§4.3.2–4.3.3: Evaluation uses only four panels in leave-one-out; L04 is a clear OOD case (roughly double life, edge impact) and already shows the worst HI and localization errors. No confidence intervals, repeated splits, or significance tests accompany MAE/RMSE/R²/SSIM. With N=4 the ~60% improvement claim is fragile. Report variability (e.g., bootstrap or multi-seed runs), temper generalization language in the abstract/conclusion, and expand the limitations discussion on specimen count and OOD behavior.","section":"§4.3.2–4.3.3, Tables 3–4, Abstract"},{"comment":"§4.3.2: OOD handling for L04 relies on an impact-region spatial embedding described only narratively (“plate partitioned into predefined spatial regions”). There is no ablation removing this prior, no formal definition of the embedding, and no sensitivity study. Because the paper attributes L04 generalization largely to this physics-informed input, the embedding must be specified (region map, encoding, how it is injected into Eq. 9 / H^(0)) and ablated against a no-region baseline so that fusion gains are not confounded with location side-information.","section":"§4.3.2, Eq. (9)"},{"comment":"§2.2.4, Eq. (12)–(13): Multitask weights λ_det=1.0, λ_loc=0.5, λ_rec=0.1 and the two-stage freeze of prediction/localization heads for 10 epochs are stated without sensitivity analysis. Given sparse, asynchronous tokens, results may depend strongly on these choices and on T_max / STFT–window tokenization. Provide at least a limited ablation on λ’s and on reconstruction loss on/off, and state how T_max and token lengths were chosen.","section":"§2.2.4, Eq. (12)–(13), §4.3.1"}],"minor_comments":[{"comment":"Abstract and §5: “nearly 60% performance improvement” should cite the exact baseline average (which metric, which folds) so the percentage is reproducible from Table 3.","section":"Abstract, §5"},{"comment":"Table 4 text: “0.0465 and 1571” appears to omit the decimal for RMSE (should be 0.1571); fix typographical error.","section":"§4.3.3"},{"comment":"Fig. 9–13 and several figure captions have spacing/OCR artifacts (“T rain”, “Data F usion”, “/s8722”); clean for production.","section":"Figures 9–13, Tables 3–4"},{"comment":"§2.1 Eq. (2): notation mixes Y_det and G_loc; later sections use detection/prediction interchangeably for HI. Unify terminology (HI prediction vs detection) throughout.","section":"§2.1–2.2"},{"comment":"Related work claims no prior PZT+FBG fusion for aircraft composites; a short check against multi-sensor ReMAP/companion papers (e.g., Broer et al. fusion diagnostics) would strengthen novelty positioning.","section":"§1"},{"comment":"Inference-time comparison (Table 5) is useful; note hardware and batch size explicitly and whether times are per sample or per panel trajectory.","section":"Table 5, §4.3.4"}],"recommendation":"major_revision","confidential_remarks":"The AE-as-label design is common in this community and does not make the ML experiment circular in the strict sense (inputs ≠ labels), but the manuscript oversells “damage” accuracy without independent NDT. For a top journal I would require the reframing and at least one external check or a much stronger limitations section; without that the contribution is solid methods-on-ReMAP rather than definitive multi-sensor damage assessment. Scope fits eess.SP / SHM venues well."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is a concrete end-to-end Transformer that tokenizes STFT patches from multi-channel guided waves and windowed FBG strain series, injects cycle-aware plus modality embeddings, masks the sparse asynchronous acquisitions, and jointly does HI regression plus pixel-wise localization with a reconstruction regularizer. That combination is new enough for the aerospace SHM literature, and the leave-one-panel-out numbers on the four ReMAP stiffened panels are cleanly better than PZT-only, FBG-only, 1D-CNN, Bi-LSTM and CNN-LSTM (MAE/RMSE under 0.1 for HI, under ~0.0465/0.1571 for localization, R^{2} still above 0.8 even on the longer-life L04 OOD case).\n\nWhat they did well: the tokenization and masking actually address the real acquisition mismatch (PZT every 5k cycles, FBG every 500), the attention maps show cross-modal focus rather than pure self-attention, and the multi-task schedule (reconstruction first, then unfreezing the heads) is sensible. The AE cumulative-energy HI is chosen after Mo–Pr–Tr screening and the localization uses anisotropic TDoA + Gaussian KDE; both are standard in this community and are not circular with the PZT/FBG inputs themselves.\n\nSoft spots, in proportion: only four panels, no error bars, free loss weights and STFT/window sizes, and no released code or data. The stress-test concern is real but overstated—AE is an independent physical modality, so the model is not fitting its own outputs; still, residual strength, stiffness drop or C-scan would have made the “true damage” claim stronger, and the elevated L04 error is consistent with label-distribution shift. Citation pattern is normal for the group and the field.\n\nThis is for people who already work on multi-sensor SHM of composites and want a reproducible architecture rather than a new theorem. It deserves a serious referee; I would send it out. Worth citing if you are building similar fusion pipelines.","headline":"Solid applied multi-modal Transformer for asynchronous PZT/FBG fusion on real ReMAP panels; gains are real relative to single-sensor and DNN baselines, but rest on AE-derived labels without independent NDT ground truth.","tokens_in":20736,"tokens_out":531,"would_cite":true,"duration_ms":6566,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A Transformer that fuses mismatched ultrasonic guided-wave and FBG strain streams predicts composite health indicators with under-0.1 error and localizes damage under fatigue loading.","keywords":["Structural health monitoring","Multisensor data fusion","Transformer network","Multitask learning","Ultrasonic guided waves","Fiber Bragg grating sensor","Health indicator","Damage localization"],"falsifier":"Repeat the identical fatigue campaign on a fifth panel instrumented with an independent imaging modality (for example, full-field digital image correlation or ultrasonic C-scan at every PZT acquisition step) and check whether the Transformer’s predicted heatmaps and health-indicator curves still match the new, non-AE labels within the reported error bounds.","tokens_in":20728,"feed_emoji":"✈️","tokens_out":962,"duration_ms":14634,"temperature":0.7,"pith_summary":"Aircraft composite panels fail gradually under cyclic loads, and no single sensor sees the full picture: ultrasonic guided waves from PZTs catch discontinuities at high frequency but only every few thousand cycles, while FBG strain gauges run more often yet lack the same spatial sensitivity. This paper claims that tokenizing both streams, padding the gaps, and letting a Transformer attend across them yields a single model that both tracks a continuous health indicator and produces a damage heatmap. On four stiffened composite skins run to failure under compression-compression fatigue, the fused model keeps health-indicator MAE and RMSE below 0.1—roughly 60 percent better than either sensor alone or ordinary deep-learning baselines—and keeps localization MAE/RMSE below 0.0465/0.1571. Attention maps show the network actually trading information between the two modalities rather than simply concatenating them. If the claim holds, condition-based maintenance of composite airframes can rest on fewer, cheaper sensor suites that still deliver both prognosis and location.","feed_headline":"Fused ultrasonic and strain sensors cut composite health error 60%","feed_subtitle":"One Transformer handles mismatched PZT and FBG streams to predict both damage growth and location under fatigue.","key_machinery":"Cycle-aware masked token fusion: STFT patches of guided-wave signals and windowed FBG strain vectors are linearly projected to a common dimension, concatenated, zero-padded to fixed length, and tagged with fatigue-cycle and modality embeddings; multi-head self-attention then performs both intra- and cross-modal fusion before shared multitask heads (health-indicator regression plus localization heatmap plus a reconstruction regularizer).","core_discovery":"When PZT spectrogram patches and multi-channel FBG time-series segments are turned into a shared token sequence, masked for missing acquisition intervals, and fused by multi-head cross-modal attention inside a Transformer, the resulting multitask network simultaneously predicts an acoustic-emission-derived health indicator and a pixel-wise damage localization map with substantially lower error than single-modality or conventional DNN baselines on the same fatigue-tested composite panels.","pith_inferences":["The same tokenization-plus-masking pattern could absorb a third sparse modality (for example, acoustic emission raw waveforms or vibration spectra) without redesigning the architecture.","Because the reconstruction head is discarded at inference, the model could be distilled into a lighter student network for on-board avionics once the fused representation is learned.","If AE labels prove biased, the attention maps themselves might be re-purposed as an unsupervised anomaly detector that does not rely on AE at all."],"forward_implications":["SHM systems can drop the requirement that every sensor fire at the same rate; missing intervals are simply masked and still contribute to both prognosis and localization.","A single trained network replaces separate pipelines for health-indicator tracking and damage imaging on stiffened composite skins.","Attention weights between PZT spectral patches and FBG channels become an inspectable diagnostic of which sensor events drive each prediction.","Cross-validation on an out-of-distribution panel (roughly double the fatigue life and a different impact site) still yields R^{2} > 0.8, suggesting the fusion transfers beyond the training lifetime distribution."],"fun_headline_variants":["Transformer fuses PZT-FBG streams to cut composite HI error 60%","Cross-modal attention predicts damage growth and location from mismatched sensors","One Transformer maps ultrasonic and strain data to HI and pixel-wise damage","Multisensor fusion via Transformer beats single-modality baselines by nearly 60%","PZT spectrograms plus FBG series yield MAE under 0.1 for composite SHM"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The paper treats averaged acoustic-emission cumulative energy and Gaussian-KDE maps built from AE time-of-arrival triangulation as the true ground-truth labels for both health-indicator regression and damage location.","fun_headline_variants_meta":{"raw":{"variants":["Transformer fuses PZT-FBG streams to cut composite HI error 60%","Cross-modal attention predicts damage growth and location from mismatched sensors","One Transformer maps ultrasonic and strain data to HI and pixel-wise damage","Multisensor fusion via Transformer beats single-modality baselines by nearly 60%","PZT spectrograms plus FBG series yield MAE under 0.1 for composite SHM"]},"model":"grok-4.5","effort":"low","cost_usd":0.004084,"raw_usage":{"total_tokens":1312,"prompt_tokens":851,"num_sources_used":0,"completion_tokens":86,"cost_in_usd_ticks":40840000,"prompt_tokens_details":{"text_tokens":851,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":375,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":851,"tokens_out":86,"duration_ms":4656,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T12:10:04.012463+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Repeat the identical fatigue campaign on a fifth panel instrumented with an independent imaging modality (for example, full-field digital image correlation or ultrasonic C-scan at every PZT acquisition step) and check whether the Transformer’s predicted heatmaps and health-indicator curves still match the new, non-AE labels within the reported error bounds.","supporting_citations":[],"review_version":1}