A report by McKinsey & Company indicates that by the year 2025 machine learning (ML) based predictive maintenance could save organizations $630 billion a year. The benefits associated with ML driven maintenance is not lost on the U.S. military. They are expected to spend billions to develop such systems. While ML does show incredible promise, applying it to military scale problems can be challenging. ML approaches popular in industries such as web services can be data hungry, requiring millions of well-balanced data points to be successful. The military and the defense industry, however, face many challenges and often do not have pristine data sets. As a result, ML algorithms must be carefully selected to effectively deal with a data set’s limitations.
This work will describe the development of a ML model that identifies performance anomalies in an aircraft subsystem. Due to the highly engineered nature of this subsystem, anomalous performance was rare. In addition, due to the subsystem’s complexity finding anomalies was challenging for a human. This made building a ML model difficult since there were few examples of anomalies and little was known about how anomalies presented themselves. To combat the limited data problem a semi-supervised ML algorithm based on Self-Organizing Maps (SOMs) was used to cluster known anomalies. Using the SOM clustering method, uncategorized examples that fell into clusters with high numbers of known anomalies were categorized as anomalous themselves. Testing results show the SOM based classifier can detect anomalous subsystem behavior with over 90% accuracy. The final paper will detail the machine learning model selection process, model development, and testing. Ultimately, this work will provide an example of how the U.S. military can apply powerful ML techniques for predictive maintenance using imperfect real-world data. © 2019 Lockheed Martin Corporation. All Rights Reserved.