Unsupervised Health-Monitoring Framework

• Unsupervised health-monitoring frameworks are methods that extract health indicators and detect anomalies from unlabeled sensor data using deep learning and clustering.
• They employ autoencoders, contrastive learning, and statistical models to learn latent representations and trigger alerts in various operating conditions.
• These approaches enhance predictive maintenance across healthcare, industrial, and structural systems by providing autonomous, interpretable monitoring with trend constraints.

Unsupervised health-monitoring frameworks are a diverse class of methodologies for inferring the state of health, predicting faults, or extracting degradation indicators from sensor data without using labeled information about faults or system state. These frameworks support prognostics, anomaly detection, and continuous surveillance across healthcare, structural monitoring, industrial asset management, and energy systems. By leveraging intrinsic system structure, causal priors about degradation, and powerful unsupervised learning algorithms—often including deep neural networks, statistical models, and graph-based representations—such frameworks enable adaptation to new modalities and unmodeled scenarios without handcrafted labels.

1. Conceptual Overview and Core Principles

Unsupervised health-monitoring frameworks aim to derive informative health indicators (HIs), detect anomalies, segment health-relevant events, or predict adverse states solely from normal operational data or time-series, without requiring direct supervision concerning fault labels or explicit degradation annotations. Conceptually, these systems operate by (i) learning normal behavior models from healthy or early-life data, (ii) applying dimensionality reduction, clustering, or self-supervised learning to extract meaningful latent representations, (iii) constructing health indicators or anomaly scores from reconstructions, representations, or statistical distances, and (iv) triggering alerts or enabling downstream analyses when deviations or trends are detected (Bajarunas et al., 2024, Hasani et al., 2017, Gabrielli et al., 5 Aug 2025, 2610.24614, Bijlani et al., 2022, Hosseini et al., 2019).

Key principles include:

• Representation learning without supervision: Using autoencoders (AEs), variational autoencoders (VAEs), LSTMs, contrastive learning, or clustering to derive latent features correlated with system health.
• Causal separation of operating conditions and degradation: Disentangling covariate effects from true degradation progression via normalization, residual modeling, or causal architectures (Bajarunas et al., 2024, Sánchez et al., 15 Jan 2026).
• Trend and monotonicity constraints: Enforcing HI temporal behavior to reflect physical irreversibility or monotonic degradation (Bajarunas et al., 2024, Perry et al., 28 Oct 2025).
• Adaptive and transfer mechanisms: Incorporating drift adaptation, domain alignment, or multi-source fusion to cope with evolving system behavior and operating environments (She et al., 2020, Michau et al., 2019).
• Ensemble modeling and multi-scale analysis: Leveraging feature fusion, signal-domain diversity, or population-wide priors to improve robustness and generalization (Perry et al., 28 Oct 2025, Soleimani-Babakamali et al., 2021).
• Autonomy and explainability: Operating without explicit expert intervention and, where possible, providing interpretable anomaly reasons or HI trends (Gabrielli et al., 5 Aug 2025, Bijlani et al., 2022).

2. Taxonomy of Methodological Approaches

Unsupervised health-monitoring frameworks can be categorized by methodological class and supported use case:

Method Class	Key Techniques	Notable Applications
Autoencoder paradigms	Sparse AE (Hasani et al., 2017), LSTM-AE (Hosseini et al., 2019, Sánchez et al., 15 Jan 2026), CAE (Bajarunas et al., 2024), DTC-VAE (Perry et al., 28 Oct 2025)	Mechanical systems, batteries, aerospace structures
Contrastive/self-supervised	Contrastive learning (with operational-time proxy) (Rombach et al., 2022), graph-based contrastive (Bijlani et al., 2022)	Asset health indicator extraction, anomaly discovery
Sequential/statistical models	HMM-FLDA event segmentation (She et al., 2020), GAN/1-Gaussian ensemble (Soleimani-Babakamali et al., 2021)	Behavioral health, structural SHM
Graph/neural anomaly frameworks	Star–graph GCN embedding, graph outlier det. (Bijlani et al., 2022)	Remote health monitoring, resource-limited contexts
Clustering and compressed sensing	Unsupervised k-means/GMM/SOM (Borthakur et al., 2018), DenStream online clustering (Hosseini et al., 2019), adaptive compressed sensing (Pagan et al., 2023)	Telecare, wearables, energy-constrained nodes
Fleet/data fusion and alignment	Incremental/adaptive HELM (Michau et al., 2019), UFAN adversarial alignment (Michau et al., 2019)	Heterogeneous industrial fleets

These methods differ in their reliance on signal feature-extraction, architectural depth, use of system-level knowledge, ability to adapt to drift, and effectiveness under sparse or highly variable operating regimes.

3. Framework Architectures and Mathematical Formulation

3.1 Representation Learning and Autoencoder-Based Systems

Autoencoders, including convolutional (Bajarunas et al., 2024), sparse (Hasani et al., 2017), and LSTM variants (Sánchez et al., 15 Jan 2026, Hosseini et al., 2019), form the basis of many frameworks. The general approach is:

1. Data normalization/conditioning: E.g., regression-based removal of operating-condition effects (Sánchez et al., 15 Jan 2026), min–max or z-score normalization, windowing.
2. Unsupervised feature learning: Training the AE to minimize the reconstruction error:

$$\mathcal{L}(X, \widehat{X}) = \frac{1}{p S} \sum_{i=1}^p \sum_{j=1}^{S} [X_{i, j} - \widehat{X}_{i, j}]^2$$

1. Health-Index extraction: The final encoder representation (or its correlation to a reference healthy state) defines the health indicator (HI) (Bajarunas et al., 2024, Hasani et al., 2017, Perry et al., 28 Oct 2025).
2. Constraint Regularization: Trendability, monotonicity, or functional constraints are introduced:
   • Negative gradient (monotonicity): $$L_{NG} = \frac{1}{m} \sum_{i=1}^{m-1} \max(0, Z_{i+1} - Z_i)$$ (Bajarunas et al., 2024)
   • Trend constraint in DTC-VAE: $$\mathcal{L}_{\rm trend} = \sum_{j=2}^N (z_j - z_{j-1} - r)^2$$ (Perry et al., 28 Oct 2025)

3.2 Clustering and Graph-Based Structures

Clustering (k-means, DenStream, GMM, SOM) is used for online event segmentation and unsupervised state discovery (Hosseini et al., 2019, Borthakur et al., 2018). For context and anomaly detection in high-dimensional time series, graph-enhanced methods deploy contextual matrix profiles (CMP) to capture temporal structure, then embed context graphs via GCNs or compute graph outlier scores (Bijlani et al., 2022).

3.3 Adaptive and Transfer Learning Mechanisms

Adaptive frameworks address temporal distribution shift and dynamic operating conditions:

• Event segmentation with HMM-FLDA: Hidden Markov Models label sessions; FLDA projects features for adaptive batch self-training, robust to drift (She et al., 2020).
• Fleet transfer via UFAN: Neural network–based adversarial feature alignment enables one-class classifiers to leverage data from heterogeneous sources (Michau et al., 2019).
• Incremental learning: Expanding the healthy baseline with low-alarm windows as new operational regimes are encountered (Michau et al., 2019).

3.4 Multi-Modal and Edge-Cloud Integration

Wearables and IoT scenarios require sensor fusion, adaptive compressed sensing to minimize transmission cost (Pagan et al., 2023), and lightweight on-device intelligence. Edge aggregation pipelines preprocess, interpolate, and package data for efficient cloud or local inference (Gabrielli et al., 5 Aug 2025).

4. Anomaly Detection, Health Index Construction, and Decision Logic

A central element of unsupervised health-monitoring is the mapping of latent model outputs to actionable scores or HIs:

• Anomaly scores are computed by future-reconstruction error (e.g., $$A(W) = \|\mathbf{X}_F - \widehat{\mathbf{X}_F}\|_2^2$$ in UniTS (Gabrielli et al., 5 Aug 2025)), AE/MLP outlier metrics (Sánchez et al., 15 Jan 2026), or synthesis-based novelty scores (GAN discriminator output) (Soleimani-Babakamali et al., 2021).
• Health indices are derived from feature correlation (AEC rate (Hasani et al., 2017)), monotonic AE latent codes (Bajarunas et al., 2024), trend-constrained VAE latents (Perry et al., 28 Oct 2025), or weighted fusion across modalities/frequencies for variance reduction (Perry et al., 28 Oct 2025).
• Thresholding employs percentile or adaptive (e.g., Peaks-over-Threshold) rules, sometimes reliability-based for minimal parameter sensitivity (Gabrielli et al., 5 Aug 2025, Soleimani-Babakamali et al., 2021).
• Early-warning logic: Confidence buffer/counting schemes (fraction of recent points outside normal clusters triggers an alarm (Hosseini et al., 2019)), trend-based change detection, or alerting based on persistent abnormality are used for robust decision-making.

5. Performance Evaluation and Deployment Considerations

Evaluation protocols are dictated by the use case, but typical metrics include:

• Correlation to ground-truth or proxy HIs: E.g., 0.97 correlation for milling machine wear (Rombach et al., 2022), 0.99–0.98 trendability for CAE-HI (Bajarunas et al., 2024).
• Balanced accuracy, F1-score, recall, specificity: For anomaly detection (88.7% balanced accuracy (Rombach et al., 2022), 0.821 F1 for stress detection (Gabrielli et al., 5 Aug 2025)).
• Recovery and generalizability: Out-of-distribution robustness in cross-mode or multi-regime tests (Bajarunas et al., 2024).
• Lead time and detection rates: Early event detection (~1.3 min before infant bradycardia with 68% AUC (Hosseini et al., 2019)).
• Scalability and real-world resource constraints: Edge inference latency (<30s per window), memory and energy footprints (single-digit ms per-chunk; <5 MB model size (Hasani et al., 2017, Gabrielli et al., 5 Aug 2025, Pagan et al., 2023)).
• Interpretability and trust: LLM-assisted anomaly explanations presented to clinicians in natural language (Gabrielli et al., 5 Aug 2025).

Case studies demonstrate versatility, e.g., successful application to railway wheels, bearings, composite structures, physiological time-series, and infrastructure SHM.

6. Domain Applications and Future Directions

Applications span:

• Industrial assets: Turbofan (CMAPSS), lithium-ion batteries, bearings, and rotary machines (Bajarunas et al., 2024, Sánchez et al., 15 Jan 2026, Hasani et al., 2017).
• Aerospace/structural health: Composite monitoring with guided waves (Perry et al., 28 Oct 2025), crack evolution in concrete (Sun et al., 6 May 2025), bridges (Soleimani-Babakamali et al., 2021).
• Healthcare and human monitoring: Remote vital sign surveillance, early event detection, ambulatory behavioral monitoring (Gabrielli et al., 5 Aug 2025, She et al., 2020, Hosseini et al., 2019, Pagan et al., 2023, Bijlani et al., 2022).
• Wearable and mobile: Activity recognition with compressed sensing on energy-constrained hardware (Pagan et al., 2023), cloud-integrated smartwatch platforms (Borthakur et al., 2018).

Emerging challenges and directions include:

• Unified architectures for multi-modal, multi-scale settings.
• Integration of explicit physical models and domain priors within representation learning (Perry et al., 28 Oct 2025, Bajarunas et al., 2024).
• Further reduction of parameter sensitivity and OOD susceptibility via causal discriminative learning and reliability-based thresholding (Soleimani-Babakamali et al., 2021, Bajarunas et al., 2024).
• Active learning and semi-supervision for faster adaptation where sparse labels are available (Perry et al., 28 Oct 2025).
• Explainability, modularity, and real-time deployment in distributed and cloud-edge environments (Gabrielli et al., 5 Aug 2025, Pagan et al., 2023).

7. Limitations, Robustness, and Generalization

Unsupervised frameworks are generally characterized by their capacity to function in absence of labeled fault data or system-specific degradation signatures, yielding robust initial models for previously unseen assets or populations. However, several limitations are acknowledged:

• Potential loss of sensitivity to subtle multi-modal or regime-dependent faults unless specialized feature sets or ensembles are constructed (Soleimani-Babakamali et al., 2021, Perry et al., 28 Oct 2025).
• Hyperparameter tuning (e.g., trend constraint rate, anomaly persistence windows) can still impact false alarm rates and detection latency.
• Fleet-wide generalization requires domain alignment strategies (e.g., UFAN) due to broad underlying system heterogeneity (Michau et al., 2019).
• **Some methods may require a minimum window of healthy data or must be retrained periodically to maintain performance under rapid non-stationarity or environmental change (She et al., 2020, Sánchez et al., 15 Jan 2026).
• Interpretability and physical meaning of learned HIs may require explicit constraints or domain knowledge to ensure monotonicity and consistency (Bajarunas et al., 2024, Perry et al., 28 Oct 2025).

Despite these, the field continues to progress toward robust, fully unsupervised, and interpretable health-monitoring solutions that are broadly adaptable to diverse sensor deployments, modalities, and operational constraints.