Zero-Shot Detection of Elastic Transient Morphology Across Physical Systems

TL;DR

This work demonstrates that a fixed latent operator, trained solely on interferometric strain transients, can perform zero-shot anomaly analysis across physically distinct domains involving elastic wave propagation. By projecting normalized time--frequency representations into a shared latent space and using a Mahalanobis-based anomaly score, the method yields monotonic drift and early warning signals in IMS--NASA run-to-failure bearings and strong separation in CWRU fault regimes, while failing to transfer to electrically dominated VSB signals. The results are supported by controlled morphology-destruction tests showing selective degradation when coherent time--frequency organization is disrupted, and by comparisons with a generic CNN baseline that reveal the importance of the interferometric morphology prior. The findings indicate that physically grounded, morphology-aware representations can enable cross-domain transfer under appropriate conditions, offering a new pathway for transferable diagnostics in wave-mediated systems. Overall, the approach provides a physically interpretable, resolution-dependent mechanism for zero-shot generalization with potential impact on cross-domain prognostics and monitoring of constrained-elastic systems.

Abstract

We test whether a representation learned from interferometric strain transients in gravitational-wave observatories can act as a frozen morphology-sensitive operator for unseen sensors, provided the target signals preserve coherent elastic transient structure. Using a neural encoder trained exclusively on non-Gaussian instrumental glitches, we perform strict zero-shot anomaly analysis on rolling-element bearings without retraining, fine-tuning, or target-domain labels. On the IMS-NASA run-to-failure dataset, the operator yields a monotonic health index HI(t) = s0.99(t)/tau normalized to an early-life reference distribution, enabling fixed false-alarm monitoring at 1-q = 1e-3 with tau = Q0.999(P0). In discrete fault regimes (CWRU), it achieves strong window-level discrimination (AUC_win about 0.90) and file-level separability approaching unity (AUC_file about 0.99). Electrically dominated vibration signals (VSB) show weak, non-selective behavior, delineating a physical boundary for transfer. Under a matched IMS controlled-split protocol, a generic EfficientNet-B0 encoder pretrained on ImageNet collapses in the intermittent regime (Lambda_tail about 2), while the interferometric operator retains strong extreme-event selectivity (Lambda_tail about 860), indicating that the effect is not a generic property of CNN features. Controlled morphology-destruction transformations selectively degrade performance despite per-window normalization, consistent with sensitivity to coherent time-frequency organization rather than marginal amplitude statistics.

Figures (6)

• Figure 1: Conceptual framework for cross-domain morphological analysis. A fixed latent operator $\mathcal{F}$, trained exclusively on interferometric strain transients (source domain), maps normalized time--frequency representations from multiple physical systems into a shared latent space $\mathcal{Z}$. The operator acts as a morphology-sensitive projection that suppresses dependence on absolute scale and sensor modality, while preserving structural organization associated with elastic wave propagation in constrained media. Zero-shot anomaly sensitivity emerges only when target-domain signals preserve compatible elastic transient structure; systems lacking mechanically mediated propagation do not exhibit separation.
• Figure 2: Zero-shot run-to-failure monitoring on IMS--NASA. Health Index trajectories for the three IMS bearing run-to-failure experiments, computed under a strict zero-shot protocol using a frozen operator trained on interferometric strain transients. The horizontal axis denotes record index (chronological order). For each record, window-level scores are aggregated via the $q=0.99$ quantile and normalized by a fixed nominal threshold $\tau$, defined as the $q=0.999$ quantile of the early-life reference distribution $\mathcal{P}_0$, corresponding to a nominal false-alarm probability of $10^{-3}$. The red dashed line marks $HI=1$ (nominal tail threshold) and the red marker indicates the first threshold crossing.
• Figure 3: Zero-shot performance and physical specificity. Receiver operating characteristic (ROC) curves for (a) the CWRU rotating machinery dataset (mechanical domain) and (b) the VSB electrical control dataset (electromagnetic domain). Strong discrimination is observed only in mechanically governed systems where elastic transient morphology is preserved, while performance collapses toward chance in electrically dominated signals, establishing a clear physical boundary for zero-shot transfer.
• Figure 4: Falsification via controlled morphological destruction. Zero-shot detection performance, quantified by window-level $\mathrm{AUC}$, under transformations that selectively degrade coherent time--frequency structure while preserving marginal amplitude statistics. Selective and transformation-dependent performance degradation indicates sensitivity to elastic transient morphology rather than to signal energy or stationary statistics.
• Figure 5: Physical interpretability of anomaly scores. Pearson correlation between anomaly scores and classical signal descriptors. Reconstruction-based scores (ConvAE) exhibit stronger coupling to amplitude- related statistics, whereas the morphological operator $\mathcal{F}$ correlates primarily with Effective Bandwidth and Spectral Entropy and remains largely orthogonal to signal energy.