Similarity-based anomaly score for fleet-based condition monitoring

An increased number of industrial assets are monitored during their daily use, producing large amounts of data. This data allows us to better monitor the health status of these asset, enabling predictive maintenance to reduce risks and costs caused by unexpected machine failure. Many condition monitoring approaches focus on assessing a machine's health status individually. Often, these approaches require historical data sets or handcrafted fault indicators. However, multiple industrial applications involve monitoring multiple similar operating machines, a fleet. By assuming the healthy behavior for the majority of the machine, deviating signatures can indicate a machine fault. In this work, we extend our previous proposed framework for fleet-based condition monitoring (Hendrickx et al.). This framework uses interpretable machine learning techniques to automatically evaluate assets within a fleet while incorporating domain knowledge if available. It is designed with four building blocks. In the first block, the user defines a similarity measure to compare machines. This measure can be both data-driven as based on domain knowledge. The second block clusters the machines based on this similarity measure. The third block assesses the health status of a machine by assigning an anomaly score where higher scores represent more deviating behavior. Finally, each of these blocks is visualized in the fourth block to guide a domain expert to set up and gain trust in the framework. The anomaly score proposed in our previous work has two shortcomings. First, its value can change very abruptly; a slight deviation can cause a machine's anomaly score to change from very low to very high. Second, the score does not accurately represent the anomalousness of a machine. A machine with the highest anomaly score is not necessarily the most deviating. Finally, the anomaly score is assigned to a group of machines. It is thus hard to assess the health status of an individual machine. As a consequence, this anomaly score offers little insights into a machine's performance. The contribution of this paper is a new implementation of the anomaly score block. Instead of basing our anomaly score on the clustering, we make use of the machine's similarities within the fleet. This solves the shortcomings of the previous anomaly score and defines an individualized, continuous scoring mechanism that represents the anomalousness of a machine.status: Published onlin