From description to code: a method to predict maintenance codes from maintainer descriptions

Abstract

Aircraft maintenance crews enter the actions performed, the time required to complete the actions, and process followed to complete the action into a system of record that may be used to support future important operational decisions such as part inventory and staffing levels. Unfortunately, the actions performed by maintainers may not align with structured, predetermined codes for such actions. This discrepancy combined with an overabundance of structured codes has led to incorrect and polluted maintenance data that cannot be used in decision making. Typically, the unstructured textual fields accurately record the maintenance action, but are inaccessible to common reporting approaches. The textual fields can be used to cleanse the structured fields, thereby making more data available to support operational decision making. This paper introduces a natural language processing pipeline to predict C-17 US Air Force maintenance codes from an unstructured, shorthand text record. This research aims to cleanse problematic structured fields for further use in operational efficiency and asset reliability measures. Novel use of text processing, extraction, clustering, and classification approaches was employed to develop a natural language processing pipeline suited to the peculiarities of short, jargon-based text. The pipeline evaluates the frequency of structured field values within the datase and selects an appropriate machine learning model to optimize the predictive accuracy. Three different predictive methods were investigated to determine an optimal approach: a Logistic Regression Classifier, a Random Forrest Classifier, and Unsupervised techniques. This pipeline predicted structured fields with an average accuracy of 93 % across the five maintenance codes.