DEEP LEARNING BASED DETECTION OF ANOMALOUS INVENTORY FLUCTUATIONS IN INDUSTRIAL SUPPLY CHAINS

Fluctuations in inventories across industrial supply chains are caused by interplay between demand, ordering, replenishment, and production which makes it challenging to detect anomalies with just one threshold or simple distributional rules. In this paper we introduce an anomaly detection approach based on four synchronized time series (inventory, demand, ordering, and replenishment). Our model captures normal operating patterns through sliding windows and computes the difference between actual and estimated inventory level as an anomaly score. We tested our model over a 240-step simulated sequence including sudden inventory drops and persistent inventory surpluses; for this sequence the multivariate model obtained an F1 score of 0.897 and an AUC of 0.951, surpassing Isolation Forest, LSTM, GRU, and baseline Transformer models. Detected anomalies matched the injected anomaly events. Results indicate potential use of multivariate time-series modeling for inventory anomaly screening, however, further validation is needed on real-world enterprise data.