This work concerns an application of a subspace-based framework with Gaussian Mixture Model (GMM) clustering for fault diagnosis in linear-parameter-varying (LPV) systems, where the scheduling variable is considered piece-wise constant and unknown. The framework relies on the classical local approach for change detection in linear time-invariant (LTI) systems, where small parametric changes are detected by tracking the expected value of an asymptotically Gaussian residual. The results for LTI systems are extended to LPV systems by clustering residuals derived from reference data at each operating point, under the assumption that all relevant operating conditions have been observed. Statistical hypothesis tests are defined for the clustered residuals, and a decision about the change is taken using a standard distance measure. We apply this method to data collected from multiple ball-bearing test-rigs, representative of the main bearing in horizontal-axis wind turbines, or small rotating machinery, like pumps, confirming the viability of the approach.
This work demonstrates that the proposed procedure can be implemented on purely sensor-based measurements, without requiring knowledge of operating conditions, like speed or load in a rotating machine.

Reliable detection of incipient faults in rotating machinery remains a critical challenge for structural health monitoring (SHM) and condition-based maintenance. This work introduces an unsupervised, data-driven framework that couples AutoRegressive modeling with Gaussian Process (GP) regression for residual analysis under varying operational conditions. The method is demonstrated on the MCC5-THU gearbox benchmark dataset [1], exploiting routinely recorded supervisory (SCADA) variables—specifically, shaft speed and torque—as exogenous inputs or scheduling parameters.

An ARX model is first trained exclusively on healthy data to describe nominal system dynamics. Residuals between predicted and measured responses are then modeled using GPs that learn the mean and variance of residuals as smooth functions of operational conditions. During inference, deviations from these GP-inferred residual distributions are statistically assessed using the Kullback–Leibler divergence, providing a probabilistic measure of anomaly.

The framework achieves robust fault sensitivity and high separability between healthy and faulty states across both speed- and torque-controlled regimes, with high classification accuracy across multiple vibration channels. Beyond its interpretability and unsupervised nature, the approach supports natural extensions to Linear Parameter-Varying (LPV) formulations, enabling context-adaptive representations of time-varying systems.

The proposed GP-augmented AR-type framework offers a lightweight yet uncertainty-aware pathway toward interpretable, data-efficient, and context-adaptive SHM for rotating machinery and related engineered systems.

[1] MCC5-THU Gearbox Benchmark Dataset: https://github.com/liuzy0708/
MCC5-THU-Gearbox-Benchmark-Datasets

Continuous monitoring of structures during operation is required for a better understanding of changes in structural parameters, for identifying problems when they occur, and for determining their causes. Several methodologies for operational modal analysis (OMA) have been developed that are capable of observing the modal properties of structures. The output-only methods of OMA assume that the forces acting on the structure are stochastic in nature, i.e., white noise.

However, these forces are not always stochastic in real life. The presence of harmonic excitations disrupts the assumption of random input, leading to false modes and incorrect modal parameters, ultimately resulting in poor system identification during operation and monitoring. The ramifications are even more severe for systems subjected to moving harmonic forces. Hence, the utilisation of a time-domain-based recursive Bayesian filtering method has been proposed, which replaces information about the harmonic forces with projected measurement output data and simultaneously identifies the system parameters.

The proposed methodology can identify the structural properties of the system while it is subjected to harmonic/moving harmonic excitation. The vibration signature of the harmonic excitation is considered to be overlapping or almost overlapping with the system's vibration signature.
The methodology has been numerically validated on a linear spring–mass–damper system and a plate model based on Reissner–Mindlin plate theory. Further it is also experimentally validated on an aluminium plate.

Unsupervised anomaly detection based on autoencoder (AE) models has become a widely adopted strategy for condition monitoring in industrial systems, particularly in scenarios where labeled fault data is scarce or unavailable. These approaches typically rely on the assumption that training and testing data follow the same distribution. However, in real-world applications, variations in operating conditions, environmental factors, sensor configurations, or machine-to-machine differences frequently induce domain shifts. Such shifts can significantly degrade detection performance and lead to false alarms, as changes in operating regimes may be misinterpreted as anomalous behavior.

While domain adaptation (DA) techniques have been extensively studied in the context of supervised fault diagnosis, their application to unsupervised anomaly detection remains limited. Most existing DA methods assume access to labeled fault data and are therefore not directly applicable to reconstruction-based detection frameworks. This work addresses this gap by proposing and evaluating two domain adaptation strategies specifically tailored for unsupervised anomaly detection using autoencoder-based models.

The first strategy is based on statistical domain alignment, where discrepancies between latent representations associated with different operating domains are minimized using covariance-based alignment techniques, such as correlation alignment. The second strategy adopts Learning Disentangled Representations (LDR) through adversarial training, explicitly decomposing the latent space into domain-invariant components, which capture the intrinsic system behavior relevant to anomaly detection, and domain-dependent components, which encode variations associated with operating conditions. In both approaches, anomaly detection is performed exclusively using reconstruction-based criteria, preserving a fully unsupervised formulation without requiring fault labels.

The proposed methodologies are evaluated on benchmark datasets comprising heterogeneous electromechanical systems operating under multiple domains, designed to emulate realistic industrial variability. Performance is assessed using standard anomaly detection metrics, with particular emphasis on the low false-alarm regime, which is critical for practical deployment in structural health monitoring and predictive maintenance applications. Results show that incorporating domain adaptation mechanisms into unsupervised AE-based frameworks significantly improves robustness to domain shifts compared to baseline models trained without adaptation. In particular, the LDR-based approach consistently achieves superior performance in systems exhibiting structured dynamic behavior, while also revealing intrinsic limitations in highly non-stationary scenarios.

Overall, this work demonstrates that domain adaptation can be effectively integrated into unsupervised anomaly detection frameworks, providing a principled pathway to enhance reliability and generalization under variable operating conditions and across multiple similar machines.