In 2018 the United States Department of Defense (DoD) released their Artificial Intelligence (AI) strategy summary. The report highlights several key AI technologies the nation needs to maintain a competitive edge. The DoD asserts that one such area, AI based predictive maintenance, is integral to ensuring equipment like aircraft and armored vehicles stay mission ready. Unfortunately, the DoD maintains varied fleets of equipment, often at lower quantities than commercial industry. These small quantities of equipment at varying operating conditions makes collecting representative data sets, often required for AI, challenging. Another factor, complicating the creation of AI for military applications, is the lack of insight into which variables best capture complex processes. This can make it challenging to determine which variables are important factors to include in an AI model.

While SMEs often provide insight into data, they may not identify the optimal combination of features. There could be bias in the SMEs recommendation, or for security reasons they may not know the true nature of the variables. As a result, another method of selecting optimal features for AI models is needed. While existing literature contains ample work on feature selection, only limited work exists dealing with small data sets. This paper describes work using Binary Particle Swarm Optimization (B-PSO) to optimize the accuracy of a Self-Organizing Maps (SOMs) based AI model for predictive maintenance trained using a small real-world data set. Testing results show that using B-PSO to select training features produces a classifier with up to 95% accuracy, 98% precision, and 72% recall. This new method increased some AI model accuracy metrics by 23% over the original baseline. The final paper will describe the novel algorithm, present testing results, and describe how this method can be used by the broader community to help increase AI model accuracy. © LMC