Deep operational normality modeling: an unsupervised framework with potential applicability to supply chain...

Area of Science:

• Machine Learning
• Supply Chain Management
• Data Science

Background:

• Modern supply chains generate vast multivariate time-series data, creating complex networks vulnerable to disruptions.
• A key challenge in anomaly detection is the scarcity of labeled anomaly data for machine learning models.
• Unsupervised learning methods are needed to model normal operational conditions and identify deviations.

Purpose of the Study:

• To propose and evaluate a novel unsupervised deep learning framework, deep operational normality modeling (DONM), for anomaly detection in sequential data.
• To address the challenge of limited labeled anomaly data by training exclusively on normal operational data.
• To demonstrate the effectiveness of DONM in identifying deviations from normal patterns for proactive risk management.

Main Methods:

• Developed DONM, an unsupervised deep learning framework utilizing a variational Transformer autoencoder (VTAE).
• Trained DONM on data representing normal operational conditions to learn temporal dependencies and probabilistic distributions.
• Utilized sequence reconstruction error (Huber loss) and Kullback-Leibler divergence to train the VTAE and generate anomaly scores.

Main Results:

• DONM demonstrated high sensitivity in anomaly detection through empirical evaluation on a real-world server machine dataset.
• The framework achieved superior performance compared to classical unsupervised methods like PCA, one-class SVM, and isolation forest.
• Key performance metrics, including AUC-ROC and F1-score, were significantly improved by DONM.

Conclusions:

• Unsupervised normality modeling using VTAEs is a highly effective paradigm for proactive risk management in sequential data.
• DONM provides a robust method for anomaly detection in complex, data-rich environments.
• The proposed method shows potential applicability to operational data in supply chains.