Revisiting Legacy Weight Relationships Using Machine Learning Techniques

This paper investigates the application of K-Means Clustering algorithms to traditional aircraft conceptual-level weight estimation techniques. As a proof of concept demonstration, application was narrowed to fuselage basic weight estimation with expansion to additional component weights as a planned follow on activity. A variety of weight sources were parsed and curated to produce a large, diverse dataset consisting of 82 separate aircraft with a corresponding new universal baseline regression to compare against. A K-Means Clustering algorithm was then employed that sorted aircraft into groupings based on configuration as well as topology and created an associated regression for each grouping. Configuration-based groupings utilized information such as a high-level abstraction of the structural layout as well as whether the aircraft is a fixed-wing or rotary-wing vehicle. Topology-cased groupings utilized information such as landing gear location and possession of a cargo ramp or wing. The configuration-based groupings showed modest improvement compared to the baseline regression while the topology-based groupings consistently outperformed both the baseline regression as well as the configuration-based regressions. Under all conditions, a subset of the data associated with fixed-wing aircraft was shown to be an outlier in regards to error as a result of a large range of weight and speed scales, as well as possible secondary pressurization impacts. Special treatment of the winged dataset led to further reduction in error based on unique design features, presenting an overall fuselage weight estimation methodology that leverages machine learning algorithms that can improve and inform existing best practices.