AI-based Continuous Learning and Human-in-the-Loop Solution for Improved Failure Diagnosis and Predictive Maintenance of Aircraft to Impact Mission Success

CBM+

Predictive maintenance solutions for the DOW can allow one to predict and prevent asset failure, enabling greater asset availability and improved mission success. Realizing these benefits requires a robust predictive maintenance solution that can maintain its accuracy over time. Traditional approaches are based on using AI to train models once before deployment but updating them with more data is a manual process. What is needed is a solution that mimics how humans continuously learn over time. A learning overtime approach would enable new failure modes to be learned and diagnosed moving forward, and new modes of normal operation to not erroneously be flagged as anomalous.

With this motivation in mind, Predictronics developed a learning over time predictive maintenance software prototype, developed through a SBIR phase I/II with the U.S Navy. The approach uses local anomaly detection models for each known normal mode of operation and failure pattern, allowing one to detect novel patterns and continuous label and learn over time. The analysis method and software prototype had promising results when evaluated on public data sets (aircraft engine, bearing and gearbox diagnosis) along with industrial data sets (marine engine, exhaust fan, CNC machine, and wind turbine health monitoring applications).

Effectively, this addresses the current gap of predictive maintenance solutions that do not maintain accuracy over time. For maintainers, they will have more robust information to act upon, driving improvements in fleet availability. Engineers can interact with this tool as well, to assist with the data review and labelling. Looking forward, we are interested in further evaluating this prototype for DOW CBM+ applications