Classification Models for RFID-Based Real-Time Detection of Process Events in the Supply Chain: An Empirical Study

RFID technology allows the collecting of fine-grained real-time information on physical processes in the supply chain that often cannot be monitored using conventional approaches. However, because of the phenomenon of false-positive reads, RFID data streams resemble noisy analog measurements rather than the desired recordings of activities within a business process. The present study investigates the use of data mining techniques for filtering and aggregating raw RFID data. We consider classifiers based on logistic regression, decision trees, and artificial neural networks using attributes derived from low-level reader data. In addition, we present a custom-made algorithm for generating decision rules using artificial attributes and an iterative training procedure. We evaluate the classifiers using a massive set of data on pallet movements collected under real-world conditions at one of the largest retailers worldwide. The results clearly indicate high classification performance of the classification models, with the rule-based classifier outperforming all others. Moreover, we show that utilizing the full spectrum of data generated by the reader hardware leads to superior performance compared with the approaches based on timestamp and antenna information proposed in prior research.