Anomaly Detection and Accuracy Measurement for Categorical Data

Abstract

The Department of Defense (DoD) recently initiated an effort to compile all inter-service maintenance data for equipment and infrastructure, requiring the consolidation of maintenance records from over 40 different data sources.  This research evaluates and improves the accuracy of this maintenance data warehouse by means of value modeling and statistical methods for anomaly detection. The first step in this work included the categorization of error-identifying metadata, which was then consolidated into a weighted scoring model. The most novel aspect of the work involved error identification processes using conditional probability combinations and likelihood measures. This analysis showed promising results, successfully identifying numerous invalid maintenance description labels through the use of conditional probability tests. This process has potential to both reduce the amount of manual labor necessary to clean the DoD maintenance data records and provide better fidelity on DoD maintenance activities.