The detection of anomalies within complex industrial systems is a critical requirement for safeguarding operational integrity, minimizing unscheduled downtime, and optimizing predictive maintenance strategies. This study proposes the Dual Autoencoder–Support Vector Data Description (DUA-SVDD) framework, an advanced unsupervised anomaly detection approach that synthesizes reconstruction-based feature evaluation with latent-space boundary modeling. By employing a dual-objective optimization scheme that integrates autoencoder-derived reconstruction loss with latent compactness constraints, the framework is designed to enhance sensitivity to both explicit feature-level deviations and implicit structural irregularities. The mathematical underpinnings of the framework are rigorously formalized, and its efficacy is systematically evaluated through controlled simulation experiments employing synthetic multivariate datasets representative of industrial monitoring environments. Empirical results demonstrate that the DUA-SVDD framework achieves robust and balanced detection performance, reliably identifying both subtle and extreme anomalies while maintaining computational efficiency. Furthermore, the inclusion of tunable hyperparameters, such as the balance weighting and percentile-based thresholding mechanisms, affords practitioners the ability to precisely calibrate detection behavior in accordance with application-specific requirements. This work offers a substantive contribution to the field of industrial anomaly detection by presenting a theoretically rigorous, operationally interpretable, and computationally viable solution. Prospective research directions include extending the framework’s application to heterogeneous real-world datasets, incorporating advanced deep learning architectures, and embedding the approach within federated learning and explainable artificial intelligence paradigms to enhance scalability, privacy, and transparency.
Cheikh et al. (Tue,) studied this question.