Zero-Shot Anomaly Analysis

• Zero-shot anomaly analysis is a technique for detecting abnormal events in data streams using pretrained models that capture time–frequency morphological features.
• The method normalizes input data, extracts latent features, and quantifies anomalies via statistical measures such as the Mahalanobis distance, achieving high discrimination in complex systems.
• It effectively transfers knowledge across domains by leveraging universal elastic-transient structures, enabling robust monitoring in systems like interferometric gravitational-wave detectors.

Zero-shot anomaly analysis denotes analytic methods that detect abnormal or rare phenomena in data streams without access to labeled anomalies or any retraining in the target domain. In this context, "zero-shot" refers to the deployment of an analysis operator that is pretrained solely on unrelated source-domain anomalies, without any fine-tuning, retraining, or adjustment using domain-specific labels in the new (target) system. The primary focus is on detecting transient, non-stationary, or highly structured deviations from the nominal background, leveraging representations that capture physical morphological features rather than only amplitude statistics. This approach has particular relevance for complex experimental systems (e.g., interferometric gravitational-wave detectors, solid-state sensors) where unanticipated noise transients ("glitches") or rare elastic-wave events severely limit detection sensitivity and reliability (Andreu, 16 Jan 2026).

1. Fundamental Concepts in Zero-Shot Anomaly Analysis

Zero-shot anomaly analysis aims to identify abnormal events in time series or sensor data, even when data from the target environment has never been directly seen or labeled by the analytic model. This contrasts with traditional supervised anomaly detection, which requires labeled fault or anomaly exemplars for each domain.

The core enabling assumption is that the morphology of true transients (e.g., burstlike, broadband, time–frequency localized events) is sufficiently generic or universal across physical systems. For example, signals generated by elastic wave propagation or mechanical impacts share characteristic spectrotemporal coherence, regardless of their source or recording sensor. Thus, a representation capturing these features in one system can act as an anomaly detector in another, provided target signals retain similar internal organization (Andreu, 16 Jan 2026).

In interferometric gravitational-wave detection, non-Gaussian strain transients ("glitches") arise from environmental, instrumental, or control-system sources and obstruct astrophysical signal search sensitivity. Advanced pipelines seek to model, filter, and veto these glitches using auxiliary channels and statistical or morphological priors, even without direct knowledge of their causal physics (Smith et al., 2011, McIver, 2012).

2. Morphology-Sensitive Representations and Cross-Domain Transfer

The cornerstone of current zero-shot anomaly analysis is the construction of a "frozen" morphology-sensitive operator—a pretrained deep neural encoding mechanism whose parameters are fixed after source-domain training and are never exposed to the target system's specific anomalies or labels (Andreu, 16 Jan 2026).

A prime example is an EfficientNet-B0 encoder pretrained on LIGO strain transients. Central workflow elements include:

• Input Normalization and Time–Frequency Representation: All target time series are resampled, transformed via STFT into fixed-size log-magnitude spectrogram windows, and standardized per window. This removes amplitude, scale, and unit dependencies, preserving only local time–frequency structure.
• Latent Feature Extraction: The neural encoder acts as $\mathcal{F}\colon X\mapsto z\in\mathbb{R}^{1280}$, mapping each normalized spectrogram window $X$ to a latent vector $z$. The weights of $\mathcal{F}$ are fixed, having been trained exclusively on interferometric glitches.
• Anomaly Quantification via Mahalanobis Distance: The anomaly score for each window is defined as

$$\mathcal{S}(z) = \sqrt{(z-\mu_0)^\top \Sigma_{\mathrm{LW}}^{-1}(z-\mu_0)},$$

where $\mu_0$ and $\Sigma_{\mathrm{LW}}$ are the mean and Ledoit-Wolf shrunk covariance of nominal ("early life") embeddings in the target environment.

• Temporal Aggregation and Health Index: Per-record anomaly quantiles $s_{0.99}(t)$ are used to define a health index

$HI(t) = \frac{s_{0.99}(t)}{\tau},$

where $\tau$ is the fixed $0.999$ quantile of the nominal distribution. Values $HI(t)>1$ signal abnormal morphological deviations, directly enabling fixed false-alarm monitoring.

Empirically, this operator achieves strong discrimination (AUC$_\text{file}\approx0.99$) between healthy and faulted regimes in mechanical vibration data—without any exposure to the target domain during training (Andreu, 16 Jan 2026).

3. Metrics for Extreme Event Selectivity and Performance

Zero-shot anomaly frameworks require robust statistical metrics to evaluate their selectivity for rare events and to control false-alarm rates. The most salient are:

• Tail Selectivity Ratio $\Lambda_{\mathrm{tail}}$:

$$\Lambda_{\mathrm{tail}} = \frac{\mathbb{P}(\mathcal{S}>\tau \mid \mathrm{abnormal})}{\mathbb{P}(\mathcal{S}>\tau \mid \mathrm{nominal})}$$

quantifies how much more frequently extreme anomaly scores are produced under faulty conditions than under normal operation. In elastic-wave dominated bearings (IMS-NASA, intermittent regime), the pretrained glitch operator yields $\Lambda_{\mathrm{tail}}\approx860$; baseline CNN features (ImageNet-pretrained) yield only $\approx2$, and electrically dominated signals yield $\approx1$.

• Area Under the Curve (AUC) Measures: Window-level and file-level AUC values quantify discrimination power at different granularities.
• Calibration via Early-Life Reference Distribution: All thresholds (e.g., $\tau$) are computed on "early-life" (nominal) data in the target system, providing fixed, transferable calibration across runs without target-domain model retraining.

The sharp selectivity decay under controlled morphological destruction (e.g., low-pass filtering, temporal smearing) and invariance under marginal amplitude statistics confirm that these metrics reflect coherent time–frequency organization, not simply power or impulsivity (Andreu, 16 Jan 2026).

4. Physical and Statistical Boundary of Zero-Shot Transfer

The efficacy of zero-shot anomaly analysis depends critically on the degree of time–frequency morphological coherence in the target signals. Transfer is highly effective for systems where anomalies manifest as broadband, time-localized elastic transients. In mechanical sensors, such as rolling-element bearings or piezoelectric vibration probes, rare fault events and impacts share the statistical and morphological properties of interferometric glitches (Andreu, 16 Jan 2026).

However, the approach fails for electrically dominated vibration signals (e.g., power-line signals, VSB), in which time–frequency structure lacks the requisite coherence, and for regimes where energy-based statistics dominate (RMS, kurtosis) but provide no discriminative substructure.

Baseline CNN representations pretrained on non-physical datasets (e.g., ImageNet) or classical reconstruction-error approaches collapse under the same evaluation protocol, indicating that morphology sensitivity is not a generic property of deep features but is encoded by domain-specific transient priors established during source-domain training (Andreu, 16 Jan 2026).

5. Relation to Traditional Anomaly Vetoes and Instrumental Data Quality

Prior to zero-shot morphological analysis, interferometric strain transient vetos were realized through hierarchical or statistical associations with auxiliary channels (Smith et al., 2011, McIver, 2012). The hierarchical veto (hveto) algorithm systematically ranks auxiliary channels by their statistical coincidence with glitch triggers, forming a minimal set of vetoes that maximizes glitch rejection while minimizing deadtime. The hveto procedure employs Poissonian significance metrics and an iterative removal protocol, achieving efficiency/deadtime ratios $>70$ in typical runs.

Alternative data quality frameworks employ category-based veto flagging, bilinear-coupling noise models, and pattern-recognition pipelines (STAMP), all requiring either explicit knowledge or statistical modeling within the target domain (Ajith et al., 2014, Collaboration et al., 2011).

Zero-shot anomaly analysis diverges in that it dispenses with auxiliary channel associations and statistical coincidences, instead projecting all target signals into a latent morphology space where the anomaly structure is measured intrinsically, permitting generalized and physically motivated detection across systems (Andreu, 16 Jan 2026).

6. Implications and Outlook

Zero-shot anomaly analysis establishes that morphology-driven latent representations, trained on physical transients in one domain, provide powerful unsupervised detectors for analogous event classes in disparate physical settings, contingent upon time–frequency structure preservation. The fixed thresholding enables operational monitoring at prescribed false alarm rates, and the selectivity metrics inform the operational envelope and limitations with respect to physical signal class.

A plausible implication is that, as physical system complexity and the diversity of fault modes increase, morphology-sensitive zero-shot operators will become increasingly valuable for commissioning, online monitoring, and root-cause analysis, particularly in domains where new or rare anomalies must be detected with minimal or no prior labeling effort. The boundary for generalization is set by the preservation of coherent elastic-transient structure; thus, domain adaptation breaks down when signal phenomenology shifts outside this class (Andreu, 16 Jan 2026).

The methodological developments in zero-shot analysis are tightly connected with advances in time–frequency normalization, neural encoding architectures, and robust statistical measures tailored to structured physical transients across scientific instruments and engineered systems.