NADA: NMF-Based Anomaly Detection in Adjacency-Matrices for Industrial Machine Log-Files

In the manufacturing and process industries, the standard approach to detecting faults or anomalies in equipment is to use condition monitoring methods. The disadvantages of these approaches are the need for expensive sensor hardware and extensive knowledge of mechanical and electronic properties. Since these challenges are not always easy to overcome, we demonstrate a novel approach that allows us to monitor the behavior of industrial equipment based on regular logs with a moderate to low number of log entries. We propose a new approach for detecting anomalies and deviations in log files using the relative or power adjacency matrix of transitions based on non-negative matrix factorization. The experiments section discusses the application to a typical manufacturing process consisting of multiple processing steps and demonstrates the ability to detect regular and irregular machine behavior.