Hybrid State Estimation and Recurrent Neural Networks for Supply Chain Disruption Detection

Abstract

Ensuring robust and resilient operational performance in modern supply chains is increasingly important as logistics systems become more digitalized, interconnected, and data-driven. Although these technological advances improve efficiency and visibility, they also introduce new vulnerabilities to supply chain fraud, inventory manipulation, procurement irregularities, and operational disruptions. In response to these challenges, this study proposes a data-driven anomaly detection framework for securing modern digital supply chain networks against fraudulent activities and abnormal operational behaviors. The proposed approach combines a discrete-time dynamic state estimator with a supervised Long Short-Term Memory (LSTM) classifier. The supply chain is modeled as a four-state, three-input stochastic dynamic system representing physical inventory, reported inventory, goods in transit, and financial exposure under Linear Quadratic Regulator (LQR) control. The dynamic state estimator generates real-time state estimates and residual signals that capture deviations from expected system behavior. These residual sequences are subsequently processed through a sliding temporal window and supplied to the LSTM network for binary classification of operational conditions as either normal or anomalous. Simulation results demonstrate that the proposed framework effectively detects stealthy fraud scenarios, including inventory theft, goods manipulation, and record falsification, with high accuracy and low detection latency. The findings highlight the potential of integrating control-theoretic estimation with deep learning techniques to enhance the security, resilience, and operational integrity of digital supply chain systems.