Unsupervised Domain Adaptation for Anomaly Detection Using Autoencoders and Disentangled Latent Representations

Abstract

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.