<p>Modern supply chains operate as complex, interconnected, and data-rich networks, generating vast amounts of multivariate time-series data. While highly efficient, this complexity renders them vulnerable to unforeseen disruptions. A significant barrier to applying modern machine learning is the inherent scarcity of labeled anomaly data. To address this, we propose deep operational normality modeling (DONM), an unsupervised deep learning framework based on a variational Transformer autoencoder (VTAE). DONM is trained exclusively on data representing normal, healthy operational conditions, learning to model the complex temporal dependencies and probabilistic distribution of this normality. The model’s objective is to minimize a combination of sequence reconstruction error (Huber loss) and Kullback–Leibler divergence, forcing it to learn a robust and structured latent representation. When presented with new sequential data, the model’s reconstruction error is used as an anomaly score; a high error signifies a significant deviation from the learned norm. We conduct a comprehensive empirical evaluation of DONM on a public, real-world benchmark dataset: the server machine dataset. Through a series of experiments, we optimize the model’s architecture and demonstrate its high sensitivity. In a final comparative analysis, DONM demonstrably outperforms selected classical and traditional unsupervised baselines—including principal component analysis, one-class SVM, and isolation forest—achieving superior performance in key metrics such as AUC-ROC and F1-score. Our findings establish that unsupervised normality modeling with VTAEs is a highly effective paradigm for proactive risk management in sequential data, demonstrating that the method is potentially applicable to supply chain operational data.</p>